Turing Tests and ChatGPT’s Sleight of Hand

 

One of the many benchmarks for AI is the “Turing test,” Alan Turing’s adaptation of the “imitation game” where an interrogator must decide which of two respondents is a computer. It is, as many have pointed out, a strangely indirect test, one that depends on the credulity of the human interrogator and the capacity of the machine to deceive (Wooldridge 2020). Will they believe the computer? And will the computer be a good enough liar? As Pantsar (2025) comments, “For the machine to pass the test, it needs to impersonate a human successfully enough to fool the interrogator. But this is puzzling in the wide context of intelligence ascriptions. Why would intelligence be connected to a form of deception?”

 

On the one hand, measuring AI through its deceptive power has the benefit of avoiding the idiocy of attempting to establish a measure of intelligence, a task deeply imbricated in racial eugenics (Bender and Hanna 2025; Wooldridge 2020). On the other, generative AI applications seem to have been developed with deception in mind–deception from all parties. The application designers want to present outputs as just as good (if not better) than the products of human work. And the humans that utilize generative AI often seek to present these outputs as their own work. And even though ethical practice demands that we acknowledge the utilization of AI, the goal is that people consuming AI material will be unable to tell the difference or to mark where the machine ends and the human begins. So even though Turing tests may be a poor measure of machine “intelligence,” they seem to fit the moment.

 

But there’s another part of the Turing Test that is especially relevant: the organization of the “imitation game” itself. Let’s go back to Turing’s 1950 article. As Sterrett points out in a series of articles, there are two games described in Turing’s landmark essay, an "Original Imitation Game” and a “Standard Turing Test” (the terms are Sterrett’s) (Sterrett 2020: 469). The first one describes two rooms, one for the interrogator, and the other for a man and woman, who communicate with the interrogator via a teletype.

In order that tones of voice may not help the interrogator the answers should be written, or better still, typewritten. The ideal arrangement is to have a teleprinter communicating between the two rooms. Alternatively the question and answers can be repeated by an intermediary. The object of the game for the third player (B) is to help the interrogator. The best strategy for her is probably to give truthful answers. She can add such things as '' I am the woman, don't listen to him! ' to her answers, but it will avail nothing as the man can make similar remarks. We now ask the question, ' What will happen when a machine takes the part of A in this game? ' Will the interrogator decide wrongly as often when the game is played like this as he does when the game is played between a man and a woman? These questions replace our original, ' Can machines think? (Turing 1950: 434)

While people have certainly looked at the gender component here, most of the attention has been on the game itself and its overall intention: deception (Patterson et al 2018). I want to focus on the whole arrangement itself.

 

The game “works” through series of limitations. The interrogator can’t go into the other room, nor can the human burst into the interrogator’s room. The interrogator is also prevented from hearing them. Communication, Turing tells us, is best accomplished through teletype. Of course, one would need such an arrangement with a computer. Yet, game rules aren’t just arbitrary, and Turing’s test tells us much about social hierarchies and workplace organization. By the time Turing wrote his test, the office had taken its contemporary form–a spatialized organizational chart made up of offices communicating with each other through a variety of technologies: mail, telephone, teletype, pneumatic tube. Inter-office communication is also rendered through a variety of intermediaries and, as “organization man” develops, cybernetic systems of communication and decision-making come to dominate managerial processes (Whyte 1956).

 

In other words, Turing’s “imitation game” rules are a description of contemporary work. Turing’s readers would have easily conjured up an image of an organization where people don’t communicate face-to-face  with each other. This has only become more marked in the intervening decades, where people are duped into long-term relationships without ever meeting their con-artist. More to the point, we are more and more called upon to judge between human- and computer outputs, a task in which humans have proven occasionally successful. But perhaps only under special conditions. 

 

One of the theses that I have suggested (both on this blog and in published work) is that the triumph of automation, algorithm and AI are initially built not upon technological change, but upon behavioral and cognitive change. Before business owners can begin to replace workers with algorithmic process and generative AI, before, in other words, those outputs can be accepted as “just as good” or “better” than human work, humans and their labor must be constrained, delimited and “de-skilled.” And more than this - everyone has to be convinced that these narrowly defined outputs are as good as we humans get. With the Turing test, human communication is reduced to lines on a teletype. With algorithmic analysis, applying to a job, getting an apartment or reading a mammogram are reduced to scores that can be ranked, patterns that can be assigned probability, etc.

 

As I explained in an earlier essay, this process of labor alienation unfolds across several steps. First, human labor is parsed out into a series of constituent functions, consistent with the Taylorization that transformed factory labor. Second, those functions are reconceptualized as algorithms: steps in a chain of operations that proceed from inputs to outputs. These might be scripts for telemarketing, decision trees for insurance claims, procedures for reporting inventory loss, etc. Next, workers are confined to those algorithmic choices, and part of the de-skilling process means penalizing workers for deviating from the script. Finally, automation (AI applications and AI-infused platforms) replaces workers. 

 

 

 

[Produced through ChatGPT]

 

The important part here is the human side of the transformation. Humanity must be reduced, and people must be convinced that algorithmic processes are interchangeable with the etiolated human. In other words, the Turing test only works if we stay in our room, if we don’t shout, if we don’t bang on the walls with our fists. And there are 2 levels of deception - one in convincing (or forcing) people to reduce themselves to narrowly defined outputs, the other in misrecognizing those algorithmic products as the total of human possibility.

 

Does generative AI involve a similar series of reductions? Of course it does. In all kinds of professions, our labor has been reduced to the production of “content”: bland and repetitive text, stereotypical images, boilerplate scripts. People produce like this in accordance with late capitalism. The only way, after all, to monetize that YouTube channel is to constantly update with new material. And that new material must fit the algorithmic desires embedded in the platform. So: you start making a lot of content, and you align it with the algorithm. The next step is, of course, to replace you with generative content. ChatGPT may have come as somewhat of a shock to us educators in 2022, but it must have come to no surprise for platform laborers, who have been human generative AI for several years now.

 

References

 

Bender, Emily and Alex Hanna (2025). The AI Con. NY: Harper.

 

Collins, Samuel Gerald (2018). Welcome to Robocracy. Anthropology of Work Review.

 

Pantsar, Markus (2024). “Intelligence is not deception.” AI & Society.

 

Patterson, W., Boboye, J., Hall, S., Hornbuckle, M. (2018). The Gender Turing Test. In: Nicholson, D. (eds) Advances in Human Factors in Cybersecurity. AHFE 2017. Advances in Intelligent Systems and Computing, vol 593. Springer, Cham. https://doi.org/10.1007/978-3-319-60585-2_26

Sterrett SG (2020) The Genius of the “Original imitation Game” test. Mind Mach 30(4):469–486. https:// doi. org/ 10. 1007/S11023-020-09543-6

Turing, Alan M.(1950). “Computing machinery and intelligence”. Mind Q. Rev. Psychol. Philos. LIX(236), 433–460.

 

Whyte, William H. (1956). The Organization Man. NY: Simon and Schuster.

 

Wooldridge, Michael (2020). A Brief History of Artificial Intelligence. NY: Flatiron Books.

 

 

Network Ghosts in the Age of Generative AI

 

What are faculty thinking about generative AI? In my role at our faculty center, I speak to faculty often on the problems they face teaching in the era of AI, and the workarounds they've come up with. The advent of publicly available generative AI platforms was not something people in my field (anthropology) or other faculty in the social sciences and humanities were clamoring for. And yet here we are. This has led to many responses: anguish, certainly, but also ways of incorporating--or at east channeling--the usage of generative AI in the classroom.

But what about faculty outside of my university? I used NodeXL to download Reddit data from the "/Professors" subreddit using the keyword "AI." This generated records of about 2500 users posting, commenting or replying for a total of 7000 contributions to the debate. I then grouped the data in clusters of similar postings, and abstracted the top words from each group as indicated by "up-vote" (which functions as more of a "like" in Reddit). As you can see, faculty were not particularly optimistic about AI in 2024. Yes, there were a couple of more computopian posters (and at least one computer scientist) who chided the community for rejecting what they saw as inevitable. But most worried that their efforts to teach writing, critical thinking, methodology and analysis were thwarted by student reliance on generative AI. Cynically, they predicted their university's tolerance for AI cheating, and speculated over their ability to continue as faculty under these conditions.

In 2024, Reddit sold their content to Google to train their large language models. This would have been been more objectionable, perhaps, if it wasn't already abundantly clear that generative AI have already been trained on Reddit, which maintains a relatively open API at a time when most social media have monetized their social network data. But what happens to that Reddit data when its re-constituted by generative AI? I decided to prompt Microsoft's Co-Pilot (to which I have enterprise-level access) to generate a spreadsheet of a Reddit conversation on AI between professors. Here's the prompt: "I would like you to generate an excel file similar to a Reddit conversation on a subreddit called "professors." The posts should discuss ChatGPT and student work from the perspective of the professor, and should include comments and replies to those comments. There should be 4 columns in the spreadsheet: A (person commenting or replying); B (person whom A is replying to); C (the text of the comment or reply); and D (the date of the reply or comment). Please populate the spreadsheet with at least 20 comments and 350 replies to those comments."

Co-pilot returned a network with with just 10 users, with 350 edges representing multiple re-postings(?) of user posts. Re-posting really isn't a thing with Reddit, so perhaps there's some confusion here with XTwitter. Since this is a much smaller network, I just labeled the 10 nodes with key words from their posts. The comments are a near "upside down" to the actual Reddit discourse over 2024, generally praising the efficiencies of generative AI and, when critical, speculating over the need for faculty at all (hence the precarity). Of course, there's a snarky comment on "Clippy," the irritating Microsoft assistant. The network itself, while smaller, is also structurally different. The actual Reddit network has a density of .001158737. In network measures of density, "1" would represent 100% connection--everyone connected to everyone else. So .0012 may not seem like much, but it's typical of social media networks where, after all, most of us don't feed the trolls and we save our replies for issues (and users) that we really care about. On the other hand, my AI-generated network has one of 0.966666667--an almost perfectly connected network where everyone has replied to everyone in a style of a polite and ploddingly inclusive panel discussion.

So, I guess that Co-Pilot does a lousy job simulating a subreddit? Yes, but, I think, more than that. It wasn't that long ago (2023) when XTwitter adopted a fee-based model for API access. That decision placed Twitter data beyond the reach of most of us. When social media data disappears behind paywalls, we (ordinary researchers) no longer really have access to the "connected action" of social media. While we can certainly look at social media, this only exposes us to our respective corners of the media platforms we inhabit and the structural components of social media are lost. But what happens when social media content is sold to OpenAI or Google Gemini? When social media disappears into a large language model, both the content and the connections are lost, and the simulated networks produced through generative AI manage to misrepresent social media on both fronts. Since Co-Pilot's inner workings are opaque to us, it is unclear if these results are the result of deliberate choice, unintended bias or something else.

Anthropology's Sad AI Archive

 

There are 3 approaches to generative AI in the classroom: 1) an outright ban on it; 2) a limited use policy that covers certain assignments or parts of assignments, and 3) an open approach that allows students to do what they would. None of these are fool-proof, whatever the intentions of the professor. Ultimately, generative AI are third-party, black-boxed products–more tempting to students, perhaps, than Wikipedia, but also more treacherous. I feel for my colleagues in the humanities attempting to wrest essays from students on Shakespeare or Aristotle: generative AI is all too good at producing a mediocre essay on these subjects. I also understand my colleagues in the computer and information sciences, who utilize these chatbots to help with their instruction.

 

But with anthropology, there are several caveats. 99.99% of writings on other peoples of the world are drenched in ethnocentrism, colonialism and racism. The internet is awash in complete nonsense about “tribes'' and their “traditional culture,” and, in generative AI, all of this is ground up and, like sausage, pumped into prompt-driven content. Yet typically, students don’t know enough to be able to distinguish a “good” and “bad” response from ChatGPT or Gemini. 

 

This is a somewhat longer way of saying that students often tried to utilize generative AI in my introductory assignments and take-home exams, and their grades suffered for it. Not because I was penalizing them for cheating; proving that they’ve used AI is almost impossible, and generative AI detectors are unreliable at best. Instead, the questions that I asked were all about the anthropology I’ve taught in classes, and generative AI is, unfortunately, only too willing to spit out all manner of palaver. Only someone who knows what to ask can minimize the racism and colonialism inherent in generative AI engines. The default is ideology. And hallucinations. 

 

One thing I want to include next year is some process of education. I really think that students don’t really know any better. The least I can do is show them that it’s not so easy and explain why that is–that generative AI is not giving them the “truth.” Or, rather, it is: the truth of colonialism and racism that underlies Western thinking about non-Western peoples. Anthropology’s sad archive. But to someone who’s never taken anthropology before, this stuff looks correct to them, and the temptation is too strong, especially in the panicked moments before a deadline.