Talk:Algorithmic bias

Wiki Education Foundation-supported course assignment

This article was the subject of a Wiki Education Foundation-supported course assignment, between 14 January 2020 and 28 April 2020. Further details are available on the course page. Student editor(s): Botchedway. Peer reviewers: Zachpiroh23.

Above undated message substituted from Template:Dashboard.wikiedu.org assignment by PrimeBOT (talk) 17:06, 17 January 2022 (UTC)

Wiki Education Foundation-supported course assignment

This article was the subject of a Wiki Education Foundation-supported course assignment, between 7 July 2020 and 14 August 2020. Further details are available on the course page. Student editor(s): Yaluys.

Above undated message substituted from Template:Dashboard.wikiedu.org assignment by PrimeBOT (talk) 17:06, 17 January 2022 (UTC)

Wiki Education Foundation-supported course assignment

This article was the subject of a Wiki Education Foundation-supported course assignment, between 24 August 2020 and 2 December 2020. Further details are available on the course page. Student editor(s): Morgan such.

Above undated message substituted from Template:Dashboard.wikiedu.org assignment by PrimeBOT (talk) 17:06, 17 January 2022 (UTC)

Stub for now, plenty more to come

Please give me some time to flesh this stub out. There are plenty of sources available on the topic, which has been widely covered in the pop press (see here, and here, and here but there are also massive reading lists compiled from a variety of high-quality academic resources. See examples here and here and here. Thanks! -- Owlsmcgee (talk) 06:58, 17 November 2017 (UTC)

Facial recognition

There's a sentence in this section that states "Software was assessed at identifying men more frequently than women..." - should this not read as "Software was assessed as identifying men more frequently than women..."? If the first version is correct, do we not need to know the results? PaleCloudedWhite (talk) 11:10, 9 December 2017 (UTC)

Good catch, PaleCloudedWhite. It's been corrected. -- Owlsmcgee (talk) 22:21, 9 December 2017 (UTC)

Thanks. PaleCloudedWhite (talk) 12:15, 10 December 2017 (UTC)

First GA review

The following discussion is closed. Please do not modify it. Subsequent comments should be made on the appropriate discussion page. No further edits should be made to this discussion.

---

GA Review

This review is transcluded from Talk:Algorithmic bias/GA1. The edit link for this section can be used to add comments to the review.

Reviewer: Farang Rak Tham (talk · contribs) 11:54, 30 May 2018 (UTC)

Introduction and limitations

Before starting this review, I'd like to state that I have little knowledge on the subject, apart from a few news reports. I do think it is a very essential subject, and will be more and more so in the near future. You will have to bear with me, because I am a newbie on this topic, but then again, for GA, you cannot be too technical, so this may turn out just well.

Overview

I have assessed the article at B now.

1. Prose:

• No copyright violations.
• The article reads well. You have made great attempts to get things explained to "dummies". Nevertheless, there are some parts left that are unclear. See detailed review below.

2. MOS:

• Remove citations in the lead which are already in the body of the article per Lead section policy, unless there are some very controversial statements in there.
• Though not required by any criteria, you might consider using shortened footnotes using the {{sfn}} template, which looks cleaner than the {{rp}} system you have used now.

3. References layout:

• There are some dead or inaccessible links in the article, which you might want to replace with other online links or archived links. Don't remove dead links though.

• There was only one dead link actually, and I've replaced it with a reference to the book where the article appears. --Owlsmcgee (talk) 23:59, 4 July 2018 (UTC)

• "us 7113917" should be expanded, in case the url dies, per WP:LINKROT.

• Fixed. --Owlsmcgee (talk) 00:04, 5 July 2018 (UTC)

4. Reliable sources:

• shirky.com, Bitch Magazine and Culture Digitally read more like blogs or essays than news coverage, and you should cite them as primary sources, using inline attribution (According to Bitch Magazine ...) and not the voice of Wikipedia. If possible, try to corroborate information from sources with data from independent, secondary, reliable sources, which will also help to show evidence of notability and relevance for the opinions quoted. You have already done this correctly with the Shirky website.
• The rest of the sources are reliable enough.

• I respectfully disagree. Bitch is a magazine with a history of editorial oversight and fact-checking, the article itself is a summary of the author's PhD thesis and dissertation, though it is a feminist magazine, that's irrelevant to the fact that it is cited in support of. Additionally, Culture Digitally is a blog run by the National Science Foundation as an explainer for scientific concepts; I don't know if I would call the National Science Foundation unreliable? And as mentioned, Clay Shirky is cited as you described. -- Owlsmcgee (talk) 00:13, 5 July 2018 (UTC)

5. Original research: None found.

6. Broadness: I believe this topic has been covered in popular culture such as here. If you can find reliable sources on this, you should add it, as it indicates how the topic is relevant for the public.

• I respect the suggestion but I don't think the popular culture section should be considered a requirement for GA status. --Owlsmcgee (talk) 00:13, 5 July 2018 (UTC)

7. Focus: Yes.

8. Neutral: Yes.

9. Stable: article is stable.

10-11. Pics: Relevant and tagged. Nicely done.

Detailed review per section

I will continue with a detailed review per section. Feel free to insert replies or inquiries.

Lead

• ... (such as a website or app) ... Isn't application better for written language?
  • Fixed and reworded. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
    • There is one instance of the word left.--Farang Rak Tham (Talk) 20:32, 8 June 2018 (UTC)
• ... even within a single use-case ...and ... even between users of the same service ... seem to contradict.
  • Agree! Clarified. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)

Methods

• ... uncertainty bias ... Isn't there somewhere this can wikilink to?
  • I could not find an article about this specific bias in algorithms. There's simply not a lot of articles about computational bias right now. We could wikilink it in red, but right now, links to uncertainty and other similar articles aren't appropriate to this very narrow idea. So, I'm passing on this suggestion for now. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
    • Okay.--Farang Rak Tham (Talk) 20:32, 8 June 2018 (UTC)
• ... algorithms may be flawed in ways that reveal personal information ... Is this also part of the definition of algorithmic bias? It doesn't sound like bias to me; more like sloppiness. Am I missing something here?
  • You're correct, I was unclear. The issue is the stereotyping in marketing, shadow profiling, and other problems. Certainly data exposure is a problem, but that isn't algorithmic bias. I've fixed it to reflect the bias-associated problem rather than privacy ones. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)

Early critiques

This section is very difficult to read, and needs to be rewritten almost completely:

• • I've reorganized it a lot and tried to clarify all your points. Hopefully it's clearer now. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
• ... human-derived operations ...: you mean procedures and working methods as in real life?
  • I mean the models that humans created to solve math problems, basically - literally the order of operations. I've tried to make it a little clearer by explaining programs as a series of steps, etc. Let me know if it reads better now. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
• ... are therefore understood to "embody law". What does this mean?
  • I've tried to make it clear. They embody law in that they take a series of steps and repeat them across all inputs, becoming a "law" for how the program executes and never changing how it executes, regardless of what data is inserted into it. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
• ... computer programs changing perceptions of machines from transferring power to transferring information How do computer programs change the perception of machines? Or do you mean people's perceptions of machines? Transferring power or information to people?
  • Good catch, that was pretty sloppy writing. Hopefully it's clearer now. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
• ... if users interpret data in intuitive ways that cannot be formally communicated to, or from, a machine. For example?
  • I think this is clear without an example, and because Weinstein was theorizing at the time, it would be WP:OR to include a contemporary example if he didn't include them himself. I hope the examples given later in the article illustrate these concepts enough for the reader. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
• Weizenbaum stated that all data fed to a machine must reflect "human decisionmaking processes" which have been translated into rules for the computer to follow. So, data must reflect processes, to translate into rules... So the rules are part of the data?
  • I've seperated his critique into two sections: Programs can be biased, and data can be biased. I hope that clears up the confusion. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
• ... and as a result, computer simulations can be built ... As a result of imagining that world incompletely?
• ... the results of such decisions ... Whose decisions?
  • This got deleted in the rewrite, if the question lingers, let me know! -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
• ... why the decision was made ... You mean, why the tourist made the decision?
• In what way are the coin tosses "correct"?
  • They aren't, which is Weizenbaum's point - I've tried to clarify it a bit. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)

Impressive rewrite. Excellent.--Farang Rak Tham (Talk) 20:32, 8 June 2018 (UTC)

Contemporary critiques

• ... natural results of the program's output ... How can the code of a program be the result of a program?
  • Good catch, this was a needlessly complex sentence. Rewritten, should be clearer. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
• These biases can create new patterns of behavior ..., Biases may also impact how society shapes itself ... For example?
  • I love examples, but again, a lot of the critiques are kind of "theorizing" and I think it might be WP:OR to say "here's an example of this" if nobody has specifically pointed to it as an example. If I find a third-party making that link, I would record it here, but it's not appropriate to draw that connection on my own until a reliable source says I should. :) -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
    • Okay, but please specify ... other elements of society ... or ... how society shapes .... WP:NOTOR may be helpful in this process. If you can't specify it, you need to find more sources. If all your secondary sources are very technical and niche-specific, try using tertiary sources like study books or encyclopedias instead. Or use popular sources written by scholars in the field--the "lay versions" of scholarly articles can often be found in news papers and popular magazines. Some of these are useful, especially when written by experts in the field. If you are unsure about the level of editorial oversight, just mention the source inline and you should be okay. Especially if those sources are written by scholars.--Farang Rak Tham (Talk) 20:32, 8 June 2018 (UTC)

I've gone ahead and summarized the research from the paper cited, hopefully this serves as an example that clarifies the meaning. -- Owlsmcgee (talk) 21:47, 14 June 2018 (UTC)

• ... weighed more heavily ... You mean, people give more authority to decisions by algorithms? Or, perceive such decisions to be more authoritative?
• Fixed. A little redundant to the next quote, but I think it's useful to reiterate. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
  • Indeed.--Farang Rak Tham (Talk) 20:32, 8 June 2018 (UTC)
• ... language frames ... I've wikilinked this now. Is my interpretation correct?
  • It wasn't, I was being too academic. I just meant the way the media or social media sites come up with language to describe stuff. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
• Sociologist Scott Lash has critiqued algorithms ... Expand or wikilink important terms.
  • I tried explaining in text. There is no article for generative power, so I hope it makes sense in context. Let me know if it's still unclear. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
    • ... they are a virtual means of generating actual ends ... is still cryptic, but the closing sentence clarifies it.--Farang Rak Tham (Talk) 20:32, 8 June 2018 (UTC)

Pre-existing

• Such ideas may reflect ... You mean the bias may reflect, right? You would not expect an institutional bias to reflect a personal bias. Ideas sounds like you are referring to the ideologies rather than the bias.
  • Yes, thanks. I've adjusted the language. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
    • The first sentence is much clearer. But I am still uncertain what this means: ... who also carry sets social, institutional, and cultural assumptions ...

* In a critical view ... Whose critical view?

• • That was a needlessly confusing addition to the sentence, and I've cut it. -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)

The example helps to explain the main points.

• • Thanks! -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)
• By attempting to appropriately articulate this logic into an algorithmic process, the BNAP inscribed the logic of the British Nationality Act into its algorithm. This sentence seems to stay the same thing twice. Am I right?
  • Not quite. The distinction is there: they were trying to reproduce the logic of immigration law into the algorithm, and in doing so, inscribed the BN act into the software. Is that clearer? -- Owlsmcgee (talk) 21:38, 14 June 2018 (UTC)
    • It is clearer now, yes.--Farang Rak Tham (Talk) 12:56, 15 June 2018 (UTC)

Detection

I have moved this section to the talk page per MOS:EMBED.--Farang Rak Tham (Talk) 21:08, 8 June 2018 (UTC)

Technical

• Flaws in random number generation can also introduce bias into results. How?
• ... what takes place beyond the camera's field of vision. Raises questions. What could be there then?
  • Tackled these both in the section. -- Owlsmcgee (talk) 22:02, 14 June 2018 (UTC)
    • Scary...--Farang Rak Tham (Talk) 12:56, 15 June 2018 (UTC)

Emergent bias

• drug breakthrough is unusual English, unless you are referring to Breakthrough therapy. Development of a new drug?
• ... clear outlines of authorship or personal responsibility ... ... for that exclusion process?

• ... a "lead member's" ... How is it determined who is the lead?

• ... it removed distant locations from their partner's preferences ... Consider cutting this out to simplify.
• {{tq|... high-rated schools ...}} You mean the schools the lead member preferred? High-rated could refer to many kinds of ratings.
  • Tackled all of these, I think it makes more sense now and is also shorter. -- Owlsmcgee (talk) 22:07, 14 June 2018 (UTC)

Correlations

• ... compared to each other in practice. Can in practice be removed?
• ... By "discrimination" against ... Confusing and redundant. I would just write By responding to or By selecting
• ... correlations can be inferred for reasons beyond the algorithm's ability to understand them. "The algorithm draws conclusions from correlations, without being able to understand those correlations?"
• ... hospitals typically give ... "hospitals without such a triage program"?

Adjusted this section too. --Owlsmcgee (talk) 22:12, 14 June 2018 (UTC)

Unanticipated uses

• ... machines may demand ... machines may expect?
• Also, certain metaphors ... For example, the British National Act Program ... How is this a metaphor?
• How is this example of British citizenship an unanticipated audience?
• Does an ATM have algorithms?

I see what you mean. I've seperated the two topics within this section, as they were being read as extensions of each other. They are two types of unanticipated users, and I hope that's clearer. The ATM was an example illustrating the concept, it wasn't meant to be about ATMs, I've removed it because I see how that's confusing. --Owlsmcgee (talk) 22:23, 14 June 2018 (UTC)

Feedback loops

• The simulation showed that public reports of crime could rise based on the sight of increased police activity, and could be interpreted by the software in modeling predictions of crime, and to encourage a further increase in police presence within the same neighborhoods. Do you mean: "The simulation discovered a correlation between increased reports of crime and increased reports of police activity"? If that is what you mean, why did the simulation encourage more police in black neighborhoods?
  • I don't mean that - I mean that reports of crime were often reported because people saw police cars. Weird, but true. I've tried to clarify. --Owlsmcgee (talk) 22:25, 14 June 2018 (UTC)
    • And successfully so. But you have not explained yet what ... the study ... refers to.--Farang Rak Tham (Talk) 12:56, 15 June 2018 (UTC)

• • Changed "the study" to "the simulation" so it's clear that the sentence refers to the same thing as the rest of the paragraph. --Owlsmcgee (talk) 00:18, 5 July 2018 (UTC)

Examples

I recommend integrating the examples into the sections about the different kinds of bias. It will help to improve understanding those sections, and will make the narrative of the body more smooth and less repetitive.

• That seems like a stylistic choice rather than one that is required for a GA. It would require an extensive rewrite of how the article is organized, which would jeopardize all progress made so far in the article toward GA status. As it stands, I feel the examples are relevant enough. -- Owlsmcgee (talk) 00:20, 5 July 2018 (UTC)

Voting behavior

• A randomized trial of Facebook users showing an increased effect of 340,000 votes among users ... This number has little meaning without context. You need to provide percentages or some other relative indication. Secondly, it isn't clear whether the friends of the users also saw the pro-voting messages. Thirdly, pro-voting is an unusual term. Maybe write which encouraged voting or something like that.

• The percentage is listed in the first line of the paragraph (a 20% swing). I let the raw number stand as that's how many additional votes resulted - I don't know how that isn't clear, given that we mentioned the percentage just two sentences earlier. I did change "pro-voting" to "which encouraged voting," thank you for the suggestion. --Owlsmcgee (talk) 00:25, 5 July 2018 (UTC)

Gender discrimination

• In fairness to Target, you might want to write they later adjusted their policies. I have read that in another book by Duhigg published in 2013.

• Could you add this info? I can't find anything online that supports it. -- Owlsmcgee (talk) 00:52, 5 July 2018 (UTC)

• This bias extends to the search ... Sentence too long, better split.

• Fixed. --Owlsmcgee (talk) 00:52, 5 July 2018 (UTC)

• ... that a suspect or prisoner will repeat a crime. A suspect has not been proven to have committed a crime. Perpetrator?

  You're right, at the point of sentencing they are still prisoners, I've adjusted the language accordingly. --Owlsmcgee (talk) 00:52, 5 July 2018 (UTC)

Sexual discrimination

• Change app into application, as indicated above.
• ... sex-offender lookup apps ... You mean, applications with blacklists of sex-offenders?
• ... saw 57,000 books de-listed ... Delisted or put on a blacklist?

  Clarified all above points. "Delisted" remains because they were delisted, not blacklisted. --Owlsmcgee (talk) 00:57, 5 July 2018 (UTC)

Lack of transparency

• Commercial algorithms are proprietary, and may be treated as trade secrets.[9]:2[16]:7[35]:183 This protects companies ... Trade secrets do not protect companies, laws about trade secrets protect companies.
• This protects companies, such as a search engine, in cases where a transparent algorithm for ranking results would reveal techniques for manipulating the service. Simplify.
• It can also be used ... The law can also be abused...
• The closed nature of the code ... The companies are closed, the code is hidden: "The closed nature of the companies..."
• ... as a certain degree of obscurity is protected by the complexity of contemporary programs ... Move the part about complexity to the next paragraph.

• All fixed, some sections moved to "complexity." --Owlsmcgee (talk) 01:29, 5 July 2018 (UTC)

Complexity

• ... large teams of programmers ... You mean, "programmers within large teams ..."?
• ... sprawling algorithmic processes ... Wikilink or define inline what this means, or simplify it.

• All fixed in the combining of text from the transparency section. --Owlsmcgee (talk) 01:29, 5 July 2018 (UTC)

Lack of data about sensitive categories

• When sources refer explicitly to US law, please say so explicitly. Same holds fro EU law, for that matter.
  • This is a general section about legal issues, when laws are mentioned I have categorized them as EU. Otherwise the section relates to various legal concerns under various systems and does not refer to any specific country's case law. --Owlsmcgee (talk) 01:36, 5 July 2018 (UTC)

• A significant barrier to understanding tackling bias in practice is that categories, such as demographics of individuals protected by anti-discrimination law, are often not explicitly held by those collecting and processing data. What does holding a category mean?

• Seemed like an editing scar, I clarified the sentence. --Owlsmcgee (talk) 01:36, 5 July 2018 (UTC)

• ... insurance rates based on historical data of car accidents which may overlap with residential clusters of ethnic minorities. Please explain further. How does this overlap? Which is false correlation?

• The point is that they do not relate to each other, that's why it's a false correlation. I've specified that the case study presents a coincidental overlap. --Owlsmcgee (talk) 01:36, 5 July 2018 (UTC)

Rapid pace of change

• ... can confuse attempts to understand them ... To understand what?

• The algorithm. I've repeated it for clarity. --Owlsmcgee (talk) 01:39, 5 July 2018 (UTC)

• ... segmenting the experience of an algorithm between users, or among the same users Simplify.

• Fixed. --Owlsmcgee (talk) 01:39, 5 July 2018 (UTC)

Rapid pace of dissemination

Please merge with section Rapid Pace of change, or expand this subsection. Normally, a GA article should not have one-line paragraphs or sections.

Europe

• ... non-binding recital ... Wikilink from the first mention.
• Why do you keep using alleged throughout this section, when the law's content is already known, as it can be quoted from?
• While these ... These regulations?
• While these are commonly considered to be new ... Too long, split. And original and originating in the same sentence is awkward.
• What do you mean by carve-outs in this context?
• The GDPR does address ... Implies a contradiction or nuance. Or do you mean "The GDPR addresses..."

• It has been argued that ... Too long, split.

• I didn't create this section but I have cleaned it up following all of your suggestions. --Owlsmcgee (talk) 01:47, 5 July 2018 (UTC)

US

• ... and uses. Uses of what?

May 2018

I will continue this detailed review once I get a response from you.--Farang Rak Tham (Talk) 13:13, 30 May 2018 (UTC)

Thank you Farang Rak Tham! I will try to respond this weekend. Do feel free to continue with your assessment if you have more to say - I tend to prefer investing enormous chunks of time into tackling these problems rather than tackling them piecemeal, but of course how you review is up to you. Thank you for a thoughtful and constructive set of recommendations! -Owlsmcgee (talk) 03:15, 2 June 2018 (UTC)

June 2018

The prose in some sections still needs a lot of copy-editing work. Try using more active voice, and less passive voice. This will help.--Farang Rak Tham (Talk) 16:08, 2 June 2018 (UTC)

Putting review on hold, as seven days have passed. You have indicated you still wish to pursue this, so I will give you another seven days. If you need even more time, you should specify a deadline yourself. The article is a very relevant subject worthy of the reader's attention, so it would certainly not be a waste of time to do some copy-editing on it.--Farang Rak Tham (Talk) 11:15, 6 June 2018 (UTC)

Hello Farang Rak Tham - I did intend to tackle it on the weekend, however, some things had come up. I will take some time with the review this week, but could you extend the timeline to June 18 so I have time to address them all in a thorough manner? I know I am a bit of a lapsed editor but it's important for me to get this article to GA status as I have created it almost entirely myself from scratch, and am very invested in this outcome! I assure you I will not abandon your work and recommendations. Thanks! -Owlsmcgee (talk) 21:14, 7 June 2018 (UTC)

Okay. Hard work!--Farang Rak Tham (Talk) 21:29, 7 June 2018 (UTC)

There is still the copy editing of the text to go through and correct, I'll get to it soon, but wanted to show a good faith effort to get some work done on this article in response to your excellent review. Thank you! -- Owlsmcgee (talk) 01:22, 8 June 2018 (UTC)

This article is very intensive in terms of copy-editing. But I will continue with it because it is very interesting material.--Farang Rak Tham (Talk) 23:22, 8 June 2018 (UTC)

Phew, I finally finished the first check. Lots of subsections. I hope you will have the time to continue with it. Some sections have little comments, others have many. Let me know if you have any questions.--Farang Rak Tham (Talk) 14:50, 15 June 2018 (UTC)

Owlsmcgee, It has been over three weeks now, and beyond the deadline which you set yourself. I am giving you until Sunday to correct the prose in the article, failing which i will have to fail the article for GA.--Farang Rak Tham (Talk) 07:29, 21 June 2018 (UTC)

Farang Rak Tham Given the enormous list of corrections, there's simply no way I can get them all done by Sunday. Feel free to fail this article as a GA and I'll resubmit after tackling the list on my own time. -- Owlsmcgee (talk) 04:16, 22 June 2018 (UTC)

Owlsmcgee, okay, but if you still want to work on it this weekend, I am willing to wait a bit. We can then see how far we can get. Whatever we do now, we don't have to do later. Up to you.--Farang Rak Tham (Talk) 16:10, 22 June 2018 (UTC)

GA progress

Good Article review progress box

Criteria: 1a. prose () 1b. MoS () 2a. ref layout () 2b. cites WP:RS () 2c. no WP:OR () 2d. no WP:CV () 3a. broadness () 3b. focus () 4. neutral () 5. stable () 6a. free or tagged images () 6b. pics relevant () Note: this represents where the article stands relative to the Good Article criteria. Criteria marked are unassessed

The discussion above is closed. Please do not modify it. Subsequent comments should be made on the appropriate discussion page. No further edits should be made to this discussion.

Moved section Detection

Cobanyastigi, I have removed this content, because it needs to be integrated as prose into the article. Simply stated, writing a Wikipedia article is not preparing a Powerpoint presentation: you need to write, not just list things. Please read MOS:EMBED first, and while you are at it, study WP:RS as well.--Farang Rak Tham (Talk) 21:02, 8 June 2018 (UTC)

There are several attempts to create methods and tools to detect and observe interacting biases in the training data:

• Google:
• Microsoft: bias-dashboard,
• Facebook: fairness-flow,
• Pymetrics: audit-ai (open source),
• Aequitas: Bias and Fairness Audit Toolkit (open source),
• IBM:

References

Second GA nomination

Owlsmcgee, currently the GA nomination has already been failed. However, i see you have been working very hard to improve the article. If you would like to continue the GA nomination, you could consider nominating the article again for GA, and i can review, if you'd like me to.--Farang Rak Tham (Talk) 05:58, 5 July 2018 (UTC)

I have re-submitted for GA, if you're happy to review it based on the feedback and responses here, I'd be happy for your efforts! Thanks. --Owlsmcgee (talk) 21:11, 6 July 2018 (UTC)

The following discussion is closed. Please do not modify it. Subsequent comments should be made on the appropriate discussion page. No further edits should be made to this discussion.

---

GA Review

This review is transcluded from Talk:Algorithmic bias/GA2. The edit link for this section can be used to add comments to the review.

Reviewer: Farang Rak Tham (talk · contribs) 16:15, 10 July 2018 (UTC)

I'll be doing this. 16:15, 10 July 2018 (UTC)

Overview

In some parts, I'll refer to the first review, and won't copy all of that here.

1. Prose: See detailed review below. No copyright violations.

2. MOS:

• Remove citations in the lead which are already in the body of the article per Lead section policy, unless there are some very controversial statements in there.

• Done! --Owlsmcgee (talk) 20:13, 28 July 2018 (UTC)

3. References layout: No problems.

4. Reliable sources: Yes. Responding to your explanation given in the previous nomination, it is advisable to attempt to access Noble's original 2012 thesis, but not required per GA criteria.

5. Original research: None found.

6. Broadness: will check this later per secondary sources. Whether popular culture should be included depends on whether this is sufficiently covered by RS.

7. Focus: As stated in the previous review, the example sections should be integrated in the other sections. The article's narrative is full of examples, so a separate section with examples is certainly redundant. I do think some of the examples are useful, but can they be integrated with the impact or types sections?

• Done! Sections have been integrated into the sections on forms of bias, and early examples into the early history. --Owlsmcgee (talk) 20:13, 28 July 2018 (UTC)

8. Neutral: Yes.

9. Stable: article is stable.

10-11. Pics: Relevant and tagged.

Detailed review per section

I'll do the lead at the end, though previous comments about the lead have all been addressed by nominator.

Voting behavior

• It isn't quite clear that the 20% and the 340,000 votes is the same group of people.

• Clarified --Owlsmcgee (talk) 20:15, 28 July 2018 (UTC)

Gender discrimination

• There are updates on Google's execution of the plans. They're mentioned briefly on Noble's blog.

• I've made it more explicit. --Owlsmcgee (talk) 20:17, 28 July 2018 (UTC)

• I'll check about Target later. It's in Duhigg's 2013 book about habits, but i can't remember it well. Something about being less aggressive about showing their predictions. May not be relevant.

• If you can find it, great, but I agree it's not absolutely needed. --Owlsmcgee (talk) 20:17, 28 July 2018 (UTC)

• It's here. Andrew Pole worked for Target. Up to you whether to include it in the article.--Farang Rak Tham (Talk) 08:03, 30 July 2018 (UTC)

Racial discrimination

• Do you mean the scores from 1920-1970 are shared to judges?

• No, I mean that for the 50 years between 1920 and 1970, this score included the nationality of the criminal's father. I've clarified. --Owlsmcgee (talk) 20:23, 28 July 2018 (UTC)

Challenges

Challenge is an euphemism for problem. Reword per WP:WTW. It's fuzzy, but part of GA, I'm afraid.

• Changed the section to a more specific "obstacles to research" which is what it's about, thanks for the suggestion. --Owlsmcgee (talk) 20:37, 28 July 2018 (UTC)

Rapid pace of change

• Can this be merged with subsection Complexity?

• Definitely a good idea. I've done so. --Owlsmcgee (talk) 20:37, 28 July 2018 (UTC)

Lack of consensus about goals and criteria

• This section is not very clearly written.
  • the availability of loans from a bank ... Complex and difficult to read, please rephrase.
  • Please check the rest of the section as well.

• Though perhaps not ideal, I've just removed this section outright, as it didn't make much sense to me upon re-reading it. It seemed to be about why and how biased algorithms get deployed in the first place, but these are handled in other sections and in more coherent ways. --Owlsmcgee (talk) 20:43, 28 July 2018 (UTC)

Europe

• Has the GDPR been implemented yet? I know there have been a number of privacy laws that have been implemented in Holland recently, though I am not sure whether this is based on European law.

  • It has been. I've removed "the "to be implemented" and made it past tense. --Owlsmcgee (talk) 20:52, 28 July 2018 (UTC)

• What are heavy exceptions?

  • I didn't write this section, but it seems odd that we are talking about laws that "weren't used often due to heavy exceptions" prior to the GDPR -- seems this section should focus on what the current regulation says, not whether past policies were effective or widely used. I've removed that piece and I think it simplifies the passage significantly, while keeping its focus on the most relevant information for this article. --Owlsmcgee (talk) 20:52, 28 July 2018 (UTC)

• Please wikilink recital at the first mention.
• as such rights may fatigue individuals who are too overburdened to use them effectively. Please clarify.

• It just means that consumers don't have to complain. I've tried to simplify that language a bit. --Owlsmcgee (talk) 20:52, 28 July 2018 (UTC)

United States

• It also recommended ... Was anything done with these recommendations?

• Clarified in text that these were guidance for a strategy. Nothing was done on a regulatory scale, however, I've added a section describing the first US law on algorithms, which was a task force created in New York City. Relevant as the first of its kind. --Owlsmcgee (talk)

Third reading

Three more comments:

• could "Online Hate Speech" and "surveillance" become subsections in the section "racial discrimination"? Perhaps the header should be edited, e.g. racial and ethnic discrimination.

• Done. --Owlsmcgee (talk) 04:08, 31 July 2018 (UTC)

• The section on regulation has not been summarized in the lead yet.

• Sparse content for the lead (really only the GDPR is lead-worthy) but I've included it, just stating the fact. --Owlsmcgee (talk) 04:08, 31 July 2018 (UTC)

• I felt that the lead does not adequately explain yet why the subject is so alarming and important, and worthy of study. This should be addressed more, per MOS:INTRO.

• I disagree that this is neccesary... the article lead is not intended to convince anyone why it is important. Wikipedia is there to provide information about the topic, not persuade people to find it important. If someone is looking for an explanation of what algorithmic bias is, they can find it in the lead now, I think. If they don't think discrimination and bias is important, it's not our job to convince them. Wikipedia is neutral, above all else. -- Owlsmcgee (talk) 04:08, 31 July 2018 (UTC)

• Perhaps my words are not ideal. MOS:INTRO says The reason for a topic's noteworthiness should be established, or at least introduced, in the lead (but not by using subjective "peacock terms" such as "acclaimed" or "award-winning" or "hit").--Farang Rak Tham (Talk) 07:33, 31 July 2018 (UTC)

• Next, I will check the broadness of the article, though i don't expect many issues there.--Farang Rak Tham (Talk) 08:03, 30 July 2018 (UTC)

Broadness

• Popular culture does not appear to be mentioned much in secondary sources (just something about Minority report here), so this can be skipped for GA.

• Great, agree. --Owlsmcgee (talk) 04:49, 31 July 2018 (UTC)

• Here, here (doi:10.1215/07402775-3813015) and here interventions are mentioned to prevent or decrease algorithmic bias. You haven't mentioned this much.

• The first article linked is a summary of ethical problems posed by algorithms and an attempt at a taxonomy, which I've already created based on a more widely-cited paper. It also essentially says that algorithmic bias isn't a problem unless it has real-world consequences, which is already addressed in the article. The second outlines some attempts by Google and Microsoft, which I've added to the end of the history section. The third is interesting but isn't actually proposing mitigation, it's instead proposing an alternative form of measuring algorithmic bias; however, it is only a working paper (a proposal for research) and hasn't actually been tested. Once there are results of an actual research study instead of a proposal, I'll be eager to incorporate that into a broader section on mitigation. --Owlsmcgee (talk) 04:50, 31 July 2018 (UTC)

• Algorithms can sometimes be better than human decisions. This is not mentioned in the article.

  Great find. I've included it as the first line in the "contemporary critiques and responses" section, to help set the context. Because the article is about the problem of bias in algorithms, I don't think it is essential to delve into all of the algorithms that are unbiased, which seems a bit like listing all the horses that a car is not. But, I think for equal weight and context, this is an important perspective, so I have introduced it very early in the section, though briefly. I think it's fair. --Owlsmcgee (talk) 04:49, 31 July 2018 (UTC)

That's all: the article is comprehensive in describing the problem and its origins.--Farang Rak Tham (Talk) 08:29, 30 July 2018 (UTC)

Okay. I think the broadness of the article passes.--Farang Rak Tham (Talk) 07:35, 31 July 2018 (UTC)

July 2018

Waiting for your responses.--Farang Rak Tham (Talk) 22:35, 10 July 2018 (UTC)

Thanks Farang Rak Tham. I will tackle these soon — I know the GA process should take about a week, but I've just seen these notes now. If it's at all possible to give me until the end of July to tackle this I would appreciate it. -- Owlsmcgee (talk) 18:48, 15 July 2018 (UTC)

Okay, I'll give you more time this time round. --Farang Rak Tham (Talk) 20:58, 15 July 2018 (UTC)

Putting on hold for now.--Farang Rak Tham (Talk) 12:50, 19 July 2018 (UTC)

Up to date, Farang Rak Tham! Have a look and let me know what you think. Thanks again! --Owlsmcgee (talk) 21:06, 28 July 2018 (UTC)

I am glad you are continuing with this. I'll take another look.--Farang Rak Tham (Talk) 07:21, 30 July 2018 (UTC)

Having looked at the lead again, I think your recent edits sufficiently address the noteworthiness of the subject matter. I don't see any reason not to pass the article for GA. Congratulations! I am happy to see you made it through. Perhaps you will think I have been a rather fuzzy reviewer, but I would not want to get the article delisted at a later time by some other editor that comes along and doesn't know how much effort you put into this.

If you have time sometimes, feel free to also review one of my GA nominations at WP:GAN#REL. See you around! --Farang Rak Tham (Talk) 07:47, 31 July 2018 (UTC)

Thanks for all of the thoughtful attention, Farang Rak Tham. I appreciate your patience as well! --Owlsmcgee (talk) 06:18, 1 August 2018 (UTC)

GA progress

Good Article review progress box

Criteria: 1a. prose () 1b. MoS () 2a. ref layout () 2b. cites WP:RS () 2c. no WP:OR () 2d. no WP:CV () 3a. broadness () 3b. focus () 4. neutral () 5. stable () 6a. free or tagged images () 6b. pics relevant () Note: this represents where the article stands relative to the Good Article criteria. Criteria marked are unassessed

The discussion above is closed. Please do not modify it. Subsequent comments should be made on the appropriate discussion page. No further edits should be made to this discussion.

Recent deletion by user:‎Owlsmcgee : Lack of consensus about goals and criteria

@‎Owlsmcgee : The two paragraphs deleted in this modification illustrate new ways people are discriminated. In one case because of company internal politics (should loans be granted based on the short term profit or based on image of the company in the public eye - i.e. long term goodwill?), and in the other case because of company interests of a quasimonopolist with that of the society at large (users that have Facebok as their news source get discriminated in what news they receive by being offered news that generate more clicks rather than news that reflect society better, users that have Youtube as their opinion forming content get discriminated by being offered more and more radical videos in an attempt to keep them watching).

May be these paragraphs should instead of being deleted, be relocated in the article as emergent or unintended impact?

I believe it is always more challenging to do impartial scientific research in a politically charged environment but I agree it should not be. The issue is rather the challenge to take any corrective action on the bias when the conflict of the short term intentions of the actors with that of their own long term interests or with the better welfare of the society at large are the cause. The article seems to lack a section for "obstacles to corrective action". Cobanyastigi (talk) 08:57, 29 July 2018 (UTC)

FAC

I saw a TEDtalk on this subject! It was absolutely fascinating! It began with asking the audience, 'how many believe computer data is dependably neutral and always fair?' then he went on to spend a good half an hour talking about what amounted to a version of 'garbage in--garbage out'. His two primary examples were about a serial killer in Sweden I think--if I remember correctly--and Michelle Obama--who knew she had raised such emotional responses? His company had to prevent any mention of the two names for a few weeks--because the programmers were upset enough they were putting in data in a manner that was creating all kinds of false outcomes. I wish I felt qualified to review this for you because it seems timely and important. I just wanted to wish you well and tell you I really hope you are successful! Good luck! Jenhawk777 21:09, 19 August 2018 (UTC)

Definition

The article body rightly says "algorithmic bias describes systematic and repeatable errors that create unfair outcomes, such as privileging one arbitrary group of users over others". But the lead described algorithmic bias as due to the values of the designers. One very important thing we have learned is that algorithms can easily be biased without *any* intention on the part of the designers to introduce bias, and indeed in ways that are wholly antithetical to the values of these designers. NBeale (talk) 22:53, 27 March 2019 (UTC)

Been on a Wikibreak but came back and saw this and wanted to say: that was a great edit to a highly visible portion of the text. Thank you! Owlsmcgee (talk)

Re-adding content that I removed

Lauraham8, I had removed a portion of content under the Google Search sub-heading which listed all the Google search results for which biased content comes up. I deleted that portion because I feel that the paragraph immediately after the list sums up and gets the point across to the reader that Google's algorithm is prone influence via systemic biases. I feel that the list will become a dumpyard for every vulgar or derogatory cliche term that can be googled and thus will degrade the overall quality of the article.

However, after my deletion, you re-added the same content without any indication of why you thought that my line of thinking was wrong. I'd like to ask you to explain why you re-added the content here so that I can understnd your veiws on the topic. Sohom Datta (talk) 15:43, 11 November 2019 (UTC)

Clarification needed

I added a {{clarify}} tag next to the following text: "However, FAT has come under serious criticism itself due to dubious funding." I think this sentence is really vague and ominous-sounding, and we ought to expand it using details from the source text. But I don't have time to make the required edits right now, so I'm flagging this here for someone to address. Qzekrom (she/her • talk) 05:38, 8 September 2020 (UTC)

References

Thanks for the suggestion Qzekrom. While I didn't add the original line, I hope the sentence reflects the referenced article better. -- Owlsmcgee (talk)

Anti-LGBT Bias

I've transferred this text from Machine learning to be incorporated in the article:

In addition to racial bias in machine learning, its usage in airport security has been found to target those who do not identify within the gender binary. In 2010, the Transportation Security Administration (TSA) started to use machine learning for bomb detection scanning. However, design justice advocates have found that the algorithm for TSA machine learning technology also detect bodies that may not align with the constructs of cisnormativity. When using ML bomb detection scanning technology, TSA agents are trained to search a person based on whether they are male or female. This has shown to be harmful towards people who identify as transgender, non-binary, gender fluid, gender-non conforming, or intersex. In 2018, nonbinary trans-feminine author and design justice advocate Sasha Costanza-Chock wrote on airport security's transition to AI: "I worry that the current path of A.I. development will produce systems that erase those of us on the margins, whether intentionally or not, whether in a spectacular moment of Singularity or (far more likely) through the mundane and relentless repetition of reduction in a thousand daily interactions with A.I. systems that, increasingly, will touch every domain of our lives." Many other people have shared their stories online about coded bias as a result of the use of machine learning in airport security procedures. For example, transgender women will almost always be flagged because their genitalia does not match the scanning guidelines set by the Advanced Imaging Technology (AIT) scanner. Since the implementation of AI-based protocol in 2010, the TSA has faced backlash from queer and transgender advocacy groups, such as the National Center for Lesbian Rights, the National Center for Transgender Equality, and the Transgender Law Center. All argue that despite the TSA's commitment to unbiased security measures, AIT and machine learning are constructed based off of biased data sets that enforce a system of oppression for people who do not identify as cisgender.

References

A separate "Algorithm (social media)" article?

There's a proposal over at Talk:Algorithm to create a section or perhaps a separate article on the general-audience article of "algorithm", as opposed to a CS/math article, since it's been talked about in the media so much. I took a shot at this at User:Hirsutism/Algorithm (social media). Is this useful? Should the scope be limited to social media algorithms, or should it cover all AI algorithms?

Wikipedia already has a wealth of such articles — this article, filter bubble, algorithmic radicalization, algorithmic transparency, regulation of algorithms, right to explanation, etc. But would a central article trying to define algorithm for a general audience be useful? --Hirsutism (talk) 15:40, 5 September 2021 (UTC)

Article Review

I thought the article was very well written, but could benefit from talking more about A.I in other countries, expanding from just Europe, The United States and India. Another country to consider would be China as they have invested a lot in A.I. as well.

67.220.29.31 (talk) 00:06, 6 October 2021 (UTC)

Merger proposal

The following discussion is closed. Please do not modify it. Subsequent comments should be made in a new section. A summary of the conclusions reached follows.

Consensus not to merge. Proposer has not brought forward a rationale for the merge, and every other editor commenting here is opposed to the idea. Felix QW (talk) 21:35, 25 April 2022 (UTC)

---

Nben7070 has proposed merging Fairness (machine learning) into this page. Qzekrom (she/her • talk) 06:18, 16 February 2022 (UTC)

My own opinion is weak oppose: it's natural to have separate articles for related concepts. Currently, Algorithmic bias serves as a general, interdisciplinary overview of the subject, whereas Fairness (machine learning) talks about the technical details of fairness measures used in machine learning. In my opinion, Algorithmic bias should discuss technical solutions at a high level more, whereas Fairness (machine learning) should explain them in greater depth.

Instead, I suggest merging Fairness (machine learning) with Fairness measure, an article that discusses fairness measures used in network engineering, as there are now papers that propose adapting the fairness measures used in network engineering (as well as welfare economics) to the artificial intelligence context (e.g. Chen and Hooker 2021). Qzekrom (she/her • talk) 04:51, 21 February 2022 (UTC)

My opinion is also a weak oppose. The Algorithmic bias page is focused on bias in general, while the page Fairness (machine learning) is focused on fairness definitions, which is a largely debated topic in the literature, and techniques for mitigation bias in machine learning, another very active domain nowadays in scientific research. I think it deserves its own page, referenced here (as it is now). I am not sure also regarding Qzekrom's proposal, since Fairness (machine learning) is not entirely dedicated to fairness measures and assessment, but also to mitigation of bias. Deltasun (talk) 10:23, 3 April 2022 (UTC)

Strong oppose. The politics of machine learning needs to be even further separated from phrases like "algorithmic bias". If anything, the present article on "algorithmic bias" should be merged under "algorithmic fairness". The field of machine learning is, via "algorithmic information" just starting to recover from several decades of technically-biased criteria for model selection due to the description languages of those information criteria being statistical rather than being algorithmic. Algorithmic information clarifies what is and is not "algorithmic bias" in a rigorous manner -- but it is completely, and by definition, utterly devoid of any notion of "ought", just as is science. A scientifically true statement may be fair or unfair, just like reality. Jim Bowery (talk) 20:13, 16 April 2022 (UTC)

The discussion above is closed. Please do not modify it. Subsequent comments should be made on the appropriate discussion page. No further edits should be made to this discussion.

Vampire image

@Esmerome: I suspect this example might be vaguely disturbing to some readers. I added WP:OR because it needs a source; I would be in favor of removing, though, as I am not sure how it helps illustrate the points in the article. Caleb Stanford (talk) 03:51, 11 July 2022 (UTC)