"We need to anticipate the traffic jams of AI before they happen"

New technologies often reshape society before their consequences are understood. Iyad Rahwan explains how science fiction science can help anticipate the impacts of AI.
 

Technology is changing faster than ever. Can society keep up?

At the moment, it's struggling. There's a wellknown concept called the Collingridge dilemma that describes a structural problem in technology governance: By the time we fully understand the consequences of a technology, it has often become too entrenched to change. But when the technology is still flexible enough to shape, we don't yet understand its impact. So societies tend to react only after problems have already emerged. What interests me is how science can help us get ahead of that—not to predict the future perfectly, but to identify possible risks and behavioral shifts before technologies are deployed at scale. That’s ultimately what led to the idea of science fiction science.

Science fiction science—What does it actually mean?

At its core, it means taking the thought experiments of science fiction and turning them into real experiments. Instead of simply imagining how people might behave in future technological scenarios, we try to gather actual empirical data. The challenge, of course, is that we can’t experiment on the future directly.

So what we do is simulate plausible future situations, present them to participants, and measure their attitudes, preferences, and behaviors. Science fiction science combines speculative imagination with the rigor of behavioral research. Rather than asking which imagined future "sounds" more plausible, we investigate whether people actually behave differently when placed—psychologically—into those futures.

There's a quote often attributed to Asimov, paraphrased by Frederik Pohl: "A good science fiction story should be able to predict not the automobile, but the traffic jam." That's exactly what I'm after—the traffic jams of AI. Not the invention itself, but the problems that emerge from its widespread use.

What makes science fiction valuable for your scientific thinking?

Science fiction is different from other speculative genres in one important way: It imagines worlds changed specifically by technology. It's not fantasy or alternative history—it's speculation grounded in what is technologically possible. And because of that, it's really about the human condition. It asks how people, societies, and moral systems change when technology reshapes the world around them.

I'm particularly drawn to hard science fiction—stories that are plausible given our current scientific understanding. They feel more real, more relatable, and that makes them much more useful as a starting point for research.

How does science fiction science differ from traditional futures research or foresight studies?

The goals are similar, but the methods are quite different. Futures research typically relies on qualitative approaches: expert panels, scenario workshops, or techniques like the Delphi method where you aggregate expert opinions.

We bring experimental methods into the picture. Rather than asking experts what they think might happen, we simulate different futures, randomly assign participants to them, and measure actual behavioral outcomes. Science fiction science doesn't replace futures studies—it complements them by adding experimental evidence to the toolkit.

What form do the experiments take?

At the simplest level, you describe a scenario and ask people to imagine themselves in it. But you can go much further than that.

A good early example is the Moral Machine experiment we ran in the mid-2010s at MIT. We wanted to understand how people think about the ethical decisions autonomous vehicles might have to make. We presented users with unavoidable accident scenarios and asked them to choose: Should the car protect its passengers or the pedestrians? Save the children or the adults?

Since then, researchers have taken this approach further using virtual reality, placing participants inside simulated vehicles so they experience the situation at first hand. And it's easy to imagine going even further, with environments built like film sets, or long-duration simulations on the lines of the Mars habitat experiments conducted in remote Arctic locations.

Is generative AI expanding the possibilities here?

It is, in two main ways. First, these models are remarkably good at exploring possible futures. Because they're trained on enormous amounts of human cultural data, they can help researchers generate plausible technological scenarios to use as the basis for experiments.

Further down the line, AI agents might also serve as "synthetic participants." Models trained on human behavioral data could let researchers run large numbers of simulated experiments quickly and cheaply, before testing the most promising ideas with real people.

That said, we need to be careful. The accuracy of the simulations depends entirely on how well the models actually reflect human behavior—and, as we know, AI can reproduce biases and errors. So AI simulations are best thought of as an exploratory tool. They are useful for generating hypotheses, but not a replacement for testing with real participants.

Looking back, where could this approach have made a difference?

It's always easier to say with hindsight, but social media is a striking example. There are now mathematical models showing that increasing connectivity can push societies past a tipping point where polarization rises sharply. That research came out recently, but similar modeling could have been done before social media platforms became so popular.

If we had spotted polarization as a systemic risk of hyperconnectivity early enough, platforms might have been designed differently—to be more locally clustered, more community-oriented. We also now know that negative and violent content spreads faster online than other types of content. Those dynamics could have been studied experimentally much earlier and potentially shaped how algorithms were built.

The same goes for social media’s impact on issues like addiction and mental health, and its effects on young people. We might not have predicted everything, but we could have started gathering evidence far sooner.

What specific research projects have you developed based on this approach?

One of our projects, launched around 2020—before the generative AI wave—examined how human behavior shifts when tasks, including ethically sensitive ones, can be delegated to AI. We discovered that people are more likely to cheat when they delegate to a machine, suggesting that AI creates a psychological distance that makes it easier to justify unethical behavior. This finding has significant implications as the systems become more widespread. For instance, it points to a need for accountability mechanisms—such as tracking who instructs the AI—especially in organizational settings where responsibility matters.

Another area we're exploring focuses on AI in positions of authority—what you might call robotic bosses or AI managers. While it might be expected that people would resist taking directions from a machine, research—including our own—shows that people often follow AI instructions, even when those instructions raise ethical concerns. It seems that humans are surprisingly willing to attribute authority to machines—a finding that is not necessarily intuitive.

Do the findings point to specific design choices that could help?

They do. For delegation to AI, the key implication is accountability: You need systems that track who instructed the AI to do what, especially in organizational settings. For AI in positions of authority, it's about transparency and oversight: making sure humans remain genuinely in the loop, rather than just nominally so.

AI will increasingly manage employees, evaluate performance, and make consequential decisions, meaning that authority will become embedded in software. Early choices around transparency, accountability, and oversight can therefore have profound and lasting effects on workplace dynamics and organizational power structures.

Are there limits to the science fiction science approach?

Absolutely. Some technologies are either too unlikely or too distant—teleportation, for example. Others involve effects that unfold over many generations, such as long-term social stratification from genetic selection. Those timescales are difficult to study meaningfully.

Science fiction science works best with plausible, near-term technologies and relatively immediate effects. It also helps to focus on specific aspects rather than massively disruptive, all-encompassing changes like the elimination of aging, which would affect almost every social institution at once.

How do you guard against your own assumptions shaping the future you study?

One way to counter this is through diversity. Cultural diversity, disciplinary diversity, and cognitive diversity all expand the space of futures we can imagine. That is why we work with an internationally diverse group of researchers and host artists-in-residence. Artists often approach the same questions from entirely different angles, which helps uncover aspects that scientists might otherwise miss.

However, we must be aware that science is never value-neutral. The questions researchers choose reflect their backgrounds, priorities, and moral intuitions. That holds for all scientific fields, not just speculative research.

Your work often highlights cultural differences. Why is that important?

Our research consistently reveals both universal patterns and meaningful cultural variation. In the Moral Machine study, for example, people across cultures shared certain moral intuitions, but differed in how strongly they weighted factors such as age or social role.

These differences matter for both design and regulation. AI systems developed in one cultural context may conflict with values in another. Just as traffic systems combine universal rules with local adaptations, AI governance will likely need a similar balance.

Can your work inform policy?

It can, and it has. Our research on self-driving cars has fed into EU regulatory discussions even before autonomous vehicles are deployed at scale. More recently, our work on the delegation of tasks to AI has highlighted the need for auditability and accountability mechanisms—the kind of guardrails that regulators are now beginning to ask for.

More broadly, debates about AI tend to get stuck between extreme optimism and extreme pessimism, and both make good policymaking harder. What science fiction science can offer is something more grounded: empirical signals about where behavioral risks are emerging, which norms are shifting, and where responsibility is starting to blur. Those signals don't predict the future precisely, but they tell you where to look and what to prepare for.

At what point is evidence strong enough to justify intervention?

This is not a challenge unique to science fiction science. It’s a general problem in evidence-based policymaking. No responsible regulator would base sweeping decisions on a single, small study. What matters is the accumulation of evidence, replication, and convergence across different methods.

The advantage of science fiction science is that this accumulation can begin earlier. Instead of waiting for a technology to cause real-world harm, we can start building an evidence base in advance. Over time, as different labs run variations of similar experiments, a clearer picture emerges.

I don't believe speculative evidence should automatically trigger strict regulation. But it should inform preparedness—shaping the questions policymakers ask, the risks they monitor, and the safeguards they consider. The goal is not to overregulate the future, but to avoid being completely unprepared for it.

How has the academic community responded to your approach?

Overall, the response has been very positive. Many researchers were already doing work that fits into this approach, even if they didn't have a shared label for it. Coining the term science fiction science helped create a common identity—a rallying point that allows people to recognize each other’s work as part of a broader methodological project.

Of course, there is skepticism, especially around questions of validity. That's healthy. Science fiction science faces unique challenges, particularly because its external validity is prospective rather than immediate. We can only fully know whether our findings hold once the future arrives.

But all science faces threats to validity. What matters is being transparent about limitations and continuously improving methods. I see this field as an evolving discipline, much like behavioral science itself, which has grown more rigorous over time.

Looking ahead, where do you see your lab in the coming years?

I want our lab to keep pushing this approach forward and, more importantly, to inspire others to take it up across different disciplines and institutions. Our students, predocs, and postdocs will carry these ideas into new contexts—that's how a research culture spreads.

The broader challenge is that technological change is accelerating, and simply reacting faster will no longer be enough. We need to combine speculative imagination with better evidence to better anticipate what the future holds. Science fiction science is one way to achieve that. And if we do it well, we won’t just predict the next technology. We'll be better prepared for its metaphorical traffic jams.