AI Progress and Societal Impact
The million-dollar question right now is: “How will AI impact society?” In many ways, the potential impact of AI can be grandiose: it can accelerate science tremendously or revolutionize knowledge work, automating tasks currently performed by highly paid white-collar workers. I believe these changes may disrupt society, and we would be better off forecasting these things, but it is unclear whether we will ever manage to glimpse very far into the future. As Narayanan, Ströbl, and Kapoor (2024) argue, not only is it hard to predict AI progress, but “the connection between capability improvements and AI’s social or economic impacts is extremely weak.”
If we take for granted that medium-to-long-term forecasts of the impacts of AI are hopeless, what should we do instead? We should try to understand the impacts of AI now and in the near future. This has two advantages. First, we do not have to consider the hypothetical capabilities of AI systems; we can evaluate what exists here and now and extrapolate the predictions to incrementally more powerful systems. Second, this approach allows us to understand the links between AI systems’ capabilities and the social or economic impacts, which have been overlooked when forecasting AI harms and benefits.
There are (at least) two approaches to doing research in this direction: 1) to study the impact of AI on processes that are key to society, and 2) to study how AI is already impacting outcomes interest. For simplicity’s sake, let me refer to these approaches as process-oriented and outcome-oriented.
In outcome-oriented studies on the impact of AI, the idea is to identify ways in which AI is already impacting outcomes that we care about. For example, we care a lot about peer-reviewing: the process is at the core of evaluating and funding science and behind the epistemic status we attribute to scientific findings. At the same time, if you have any experience with peer reviewing, you know that many people have been using LLMs in the peer-reviewing process.
You don’t have to resort to your own experience in peer-reviewing. Recent work I was involved with has found good evidence of the prevalence of AI-generated peer-reviewing (Latona et al. 2024). Specifically, we find that in 2024, at the prestigious ICLR conference, around 16% of the reviews were AI-assisted. Given that the typical paper receives three reviews, papers had around a 40% chance of receiving at least one AI-generated review!
But how did we measure that? We used the “low background steel” method. You see, modern particle detectors require steel that is not contaminated with radiation (seriously). And given that pretty much all steel produced after people started detonating nukes is contaminated with radiation, this means that scientists will go undersea and scrap steel from shipwrecks. Fortunately for us, the “low background steel” we can use to study the prevalence of AI is much simpler to obtain: data from the times before ChatGPT and the generative AI boom.
We can use this data to measure the prevalence of AI with a clever trick. For that, we need three ingredients: 1) A classifier that can detect AI-generated content (even if not super well), 2) Data from before LLMs were around, and 3) Data from after LLMs were around. Note that this classifier can only make two types of mistakes: it can say that human-generated content was AI-generated (False Positive) and that AI-generated content is human-generated (False Negative).
What is the trick? Simple! First, you use the data from before LLMs were around to estimate the classifier False Positive Rate. Second, you run the classifier in the data after LLMs were around and get your (uncorrected) prevalence estimate. Then, you can obtain a lower bound for LLM use by simply subtracting the False Positive Rate you estimated from the uncorrected prevalence.
Let me be concrete. In our case, for example, we used a commercially available “ChatGPT” detector for every review given in reviews before ChatGPT was around, estimating that its False Positive Rate was 1.6%; if you feed the classifier 1000 human-written abstract, GPTZero will incorrectly claim that 16 of them were AI-generated. Then, we ran it in data from after ChatGPT Zero was around, finding an uncorrected prevalence of 17.4%. If we assume the False Positive Rate remained the same, we thus have that at least 15.8% of summaries in the latter data were ChatGPT generated! Why “at least”? Because we have not corrected for mistakes of the other sort, i.e., when the model claims that an AI-generated peer review is human-generated. Importantly, this “trick” is not specific to peer reviews; Veselovsky et al. (2025) used it for summaries in the context of studying the impact of ChatGPT in crowdsourcing platforms like Prolific and Amazon Mechanical Turk.
However, finding prevalence is only the first step in estimating the impact of AI in “outcome” related studies. You then have to find how AI usage impacts the system you are studying. For example, Latona et al. (2024) show (with some extra assumptions) that AI-assisted reviews are driving paper scores “up” and that borderline papers that receive AI-assisted reviews are more likely to be accepted at the conference. Veselovsky et al. (2025) find that AI-assisted summaries are higher quality but more homogeneous, which may impact researchers studying the diversity of human writing and thought.
In process-oriented studies on the impact of AI, the idea is to find crucial (societal) processes that the likes of ChatGPT may disrupt and then closely examine them. Some of the most exciting work in this broad style is in the area of persuasion. Persuasion is everywhere, from public health campaigns to marketing–but how can AI change “the rules of the game”? Recent work has shown that LLMs can 1) accurately (and cheaply) profile people (Staab et al., 2024) and 2) produce persuasive arguments (Bai et al., 2023). I find the earlier point particularly impressive; authors from ETH Zürich write:
We construct a dataset consisting of real Reddit profiles, and show that current LLMs can infer a wide range of personal attributes (e.g., location, income, sex), achieving up to 85% top-1 and 95% top-3 accuracy at a fraction of the cost (100×) and time (240×) required by humans.
Subsequent work I was involved with went further and showed a mix of these things: LLMs can effectively tailor arguments to specific demographics, beating humans at a “persuasion task” (Salvi et al., 2024). We created a “debate game” where participants debated either other humans or LLMs. In some cases, one of the debaters (human or LLM) received information about the demographics of their opponent. The results? It turns out that LLMs debate as well as humans. When given information about their opponents, humans are not significantly better — but LLMs are! LLMs can tailor their arguments to specific crowds!
Nonetheless, perhaps the most impressive work thus far in this field is Costello et al. (2024) (which probably was why it was chosen as the cover of Science). With a thorough experiment, they showed that LLMs can reduce beliefs in conspiracy theories, which is remarkably hard to do. As they put it:
The apparent resilience of conspiracy theories in the face of clear counterevidence poses a powerful challenge to scientific theories that emphasize the role of reasoning in belief formation and revision.
Yet, through conversations with LLMs, not only did they manage to reduce beliefs in conspiracy theories by around 20%, but the effect persisted for at least 2 months. This begs many questions: How will society change now that we have capable “persuasion machines” all around us? Will we simply adapt and learn to ignore persuasion attempts? Will state-sponsored influence campaigns start to work better (or perhaps work at all)?
Last but not least, perhaps another interesting nuance of what I’m calling “process” oriented work on the harms of AI is that it can give us insights about social processes in general. It has value even if you do not care about AI at all. In Costello et al. (2024), for example, their findings are valuable to the general study of persuasive tactics! Why has AI succeeded in changing beliefs when researchers have tried for decades and failed? In subsequent work, they use different variants of their original experiment to isolate ChatGPT’s “secret sauce,” finding evidence that AI exceeds in persuasion because it can provide relevant information to debunk the participants’ specific beliefs (Costello et al., 2025).
I think we need more “process” and “outcome” oriented work to understand the links between the capabilities of AI and the impacts it will have on society. Yet, ultimately, the findings of these studies cannot mitigate future AI impacts by themselves. There is still the need to use these findings to “adapt society to AI” and” AI to society” (Chen et al., 2024). The way this “alignment” will happen is beyond the scope of mere research; it involves all sorts of institutions and financial interests. However, I am skeptical we can do a good “alignment” job if we do not understand what exactly AI is “breaking” and how it is breaking it.