Jan K. Woike
LightNews LightNews
banner
jan-k-woike.bsky.social
Jan K. Woike
@jan-k-woike.bsky.social
660 followers 590 following 24 posts
University of Plymouth (UK); Max Planck Institute for Human Development, Berlin; Decision Making Researcher
Posts Replies Media Videos
Reposted by Jan K. Woike
olivia.science
Now, we tackle in turn the issues we outlined in the abstract:

First, the displacement of human participants, tracing the history of technologies to recruit real people, we end up forgetting the whole point: gathering — not fabricating! — human data. More at section 3: doi.org/10.31234/osf...

5/
3 Displacement of Participants Many technological solutions to recruiting and managing participants have been proposed: from original pen and paper sign up sheets and physical bulletin boards to computer systems like SONA, which digitised and automated some of these processes, and then furthermore to platforms like Amazon’s Mechanical Turk (MTurk), which allows participants to be outsourced to sweatshops (e.g. C. A. Anderson et al., 2019; Douglas et al., 2023; Gamblin et al., 2017; Gray & Suri, 2019; Hauser & Schwarz, 2016; Wagner et al., 2022). It may be the case that systems like MTurk, which is named after the infamous pseudo-automaton which beat Napoleon at chess (Manninger, 2024), indeed have clear pros and cons, and are perhaps not inherently abusive. However, they are clearly open to abuse (C. A. Anderson et al., 2019; Guest, 2024; Newman, 2019; Pettit, 2024c; Stephens, 2023). MTurk’s namesake, the original Mechanical Turk, was a device that comprised a cupboard with a chess board on top and which across the player sat an orientalist puppet. However much it may have seemed the puppet played chess, it was instead a hidden human who moved the pieces. In the present, sweatshops based on Amazon’s MTurk technology undergird all modern AI products (Alba, 2023; Bansal, 2025; Equidem, 2025; Gershgorn, 2017; Gray & Suri, 2019; Manninger, 2024; Perrigo, 2023; Stephens, 2023; Streitfeld, 2025; Suri, 2019; Yang et al., 2020). This human-in-the-loop technique has had applications in the 2010s with deep artificial neural networks that performed what is claimed to be human-like vision, and in the 2020s with reinforcement learning from human feedback (e.g. Birhane et al., 2023; Guest & Martin, 2025b; Kirk et al., 2023; Manninger, 2024; Prabhu & Birhane, 2020; Suri, 2019). The next iteration of this is to completely obfuscate the human-in-the-loop (Guest, 2025; Guest & Martin, 2025a; Shiffrin & Mitchell, 2023) and present AI products, such as those powered by large language models (LLMs), as able to stand in for, displace, human participants or their data (cf. Crockett & Messeri, 2023; Dillion et al., 2023; Jamieson et al., 2023; Rilla et al., 2025; Rossi et al., 2024; Schröder et al., 2025). This is problematic for many reasons (viz. Guest, Suarez, et al., 2025; van Rooij & Guest, 2025), but some lesser discussed angles are that the AI products cannot replace human participants but are merely a) returning unethically sourced (pre-existing) human data (Alba, 2023; Bansal, 2025; Equidem, 2025; Guest, 2025; Perrigo, 2023). And as such, b) these data have statistically unfavourable properties according to what are stated norms of the field, such as comprising double dipping (Y. Liu et al., 2024; Villalobos et al., 2024; What is most disparaging if we continue down this route, is not only that it leads to data fabrication on steroids but that also we will depend on a machine that, for lack of a less anthropomorphic term, lies, and lies about lying by design (Bender & Hanna, 2025; Bender et al., 2021; DeVrio et al., 2025; Edwards, 2023b; Hicks et al., 2024; Metz, 2023; Xu et al., 2025; Zhao et al., 2025). It deceives through the Eliza effect, the cognitive bias towards assigning human-like traits to chatbots (Borau, 2025; Dillon, 2020; Koike & Loughnan, 2021; Weizenbaum, 1966), and the BarnumForer effect, the phenomenon where one thinks generic person
October 4, 2025 at 5:58 AM
Reposted by Jan K. Woike
olivia.science
Second, if we uncritically adopt AI in psychology, we outsource programming of our experiments to companies. This undoes a lot of open source: experiments are under industry capture, written without formal specifications. It deskills us: psychologists spent the last decade learning to code.

6/
4 Outsourcing Programming to Companies [Sanford’s mechanical vernier chronoscope] had two pendulums, one set longer than the other but both set at known lengths. When the stimulus appeared, the longer pendulum was released mechanically. When the participant pressed the key in response, the shorter pendulum was similarly released. By counting the number of swings it took the shorter pendulum to catch up with the longer, experimenters could use a mathematical formula to calculate the difference in time between the two events. (Evans, 2000, p. 322–323) A century later, all this is now done by programming computers to display stimuli, and collect and analyse data. The largely harmless and perhaps even beneficial (cf. Guest, 2025) transition to computational experimentation, where all instruments (e.g. EEG, fMRI, eye-tracking) are hooked up to the codebase for running experiments and the database for storing participant data, has been completed through: the adoption of programmes such as SPSS for statistics (Landau & Everitt, 2003) and E-Prime for presenting stimuli (Spapé et al., 2019); and training psychologists to use programming languages (e.g. R, Python; British Psychological Society, 2017; Peirce et al., 2022). The last part, from the mid 2010s onwards, where psychologists spent time and effort learning how to program is a massively laudable case of reskilling (N. D. Anderson, 2016; Guest & Forbes, 2024; Scherer et al., 2019) From the 2020s, however with the rise of the kinds of AI products we critique herein, this is under threat. If scientists accept the AI “snakeoil” and “con” (Bender & Hanna, 2025; Bjarnason, 2023; Narayanan & Kapoor, 2024) and use it to deskill themselves from learning proper programming, then code reproducibility and verifiability are lost causes; not just short-lived, dead on arrival (Becker et al., 2025a). If the code is for running experiments, strange side-effects will be present from so-called ‘vibe coding’ (Harkar, 2025), which is ‘programming’ using a chatbot such that nothing is checked or verified against software engineering principles and only the output being seen as acceptable under a small sample of conditions is evaluated. This means that because standards are ignored and best practises are violated the stimuli may be presented in the wrong order, participant data may not be properly or securely saved, and the code may be unmaintainable (Burgess, 2023, 2025; El-Mhamdi et al., 2022; Greshake et al., 2023; Ming, 2025; Tangermann, 2025). And the same is true for this non-engineered code in the case of analysing data, which will not only not be analysed, but open to all sorts of bugs. To add insult to injury, because the scientist is not practising programming or never learned how to code, all this will be outside their skill level and they will be unable to detect and fix these issues (Becker As Danielle Navarro (2015) says about shortcuts through using inappropriate technology, which chatbots are, we end up digging ourselves into “a very deep hole.” She goes on to explain: The business model here is to suck you in during your student days, and then leave you dependent on their tools when you go out into the real world. [...] And you can avoid it: if you make use of packages like R that are open source and free, you never get trapped having to pay exorbitant licensing fees. (pp. 37–38) These were and are the reasons to switch from e.g. SPSS, which is owned by IBM, to R or Python (Wills, 2019). The same and more hold for being locked in and addicted to AI products, which have closed or industry-controlled source code (Hao, 2025; Liesenfeld & Dingemanse, 2024; Liesenfeld et al., 2023; Maffulli, 2023; Maris, 2025; Mirowski, 2023; Solaiman, 2023) , and mislead us into overestimating our programming abilities (Lehmann et al., 2025)
October 4, 2025 at 6:07 AM
Reposted by Jan K. Woike
olivia.science
Third, the peculiar idea that somehow we don't need to read, write, or perform literature reviews anymore; popping up like a satanic mushroom in almost all so-called OK uses of LLMs.

Companies writing our papers via their chatbots is not scientific at all. See section 5: doi.org/10.31234/osf...

7/
table 1 extract from Guest, O., & van Rooij, I. (2025, October 4). Critical Artificial Intelligence Literacy for Psychologists. https://doi.org/10.31234/osf.io/dkrgj_v1 5 Ghostwriter in the Machine A unique selling point of these systems is conversing and writing in a human-like way. This is imminently understandable, although wrong-headed, when one realises these are systems that essentially function as lossy2 content-addressable memory: when input is given, the output generated by the model is text that stochastically matches the input text. The reason text at the output looks novel is because by design the AI product performs an automated version of what is known as mosaic or patchwork plagiarism (Baždarić, 2013) — due to the nature of input masking and next token prediction, the output essentially uses similar words in similar orders to what it has been exposed to. This makes the automated flagging of plagiarism unlikely, which is also true when students or colleagues perform this type of copypaste and then thesaurus trick, and true when so-called AI plagiarism detectors falsely claim to detect AI-produced text (Edwards, 2023a). This aspect of LLM-based AI products can be seen as an automation of plagiarism and especially of the research paper mill (Guest, 2025; Guest, Suarez, et al., 2025; van Rooij, 2022): the “churn[ing] out [of] fake or poor-quality journal papers” (Sanderson, 2024; Committee on Publication Ethics, In addition, who is held accountable if nobody with intent authored the text? Because while the original data fed into the system is certainly written with goals, messages, and audiences in mind jumbling this into ad-libbed word salad removes authorial intent (Bender et al., 2021). So do the companies who own the chatbot own the text or do the original authors? These questions denote legal battles, which are being currently fought in the public eye and which affect all of us in all roles, not just as academics (Creamer, 2025; Knibbs, 2024; Reuters, 2025). Either way, even if the courts decide in the favour of companies, we should not allow these companies with vested interests to write our papers (Fisher et al., 2025), or to filter what we include in our papers. Because it is not the case that we only operate based on legal precedents, but also on our own ethical values and scientific integrity codes (ALLEA, 2023; KNAW et al., 2018), and we have a direct duty to protect, as with previous crises and in general, the literature from pollution. In other words, the same issues as in previous sections play out here, where essentially now every paper produced using chatbot output must declare a conflict of interest, since the output text can be biased in subtle or direct ways by the company who owns the bot (see Table 2). Seen in the right light — AI products understood as contentaddressable systems — we see that framing the user, the academic in this case, as the creator of the bot’s output is misplaced. The input does not cause the output in an authorial sense, much like input to a library search engine does not cause relevant articles and books to be written (Guest, 2025). The respective authors wrote those, not the search query!
October 4, 2025 at 6:16 AM