AI and the Future of Shared Knowledge

Imagine a child sitting beside a fire, listening to her grandmother explain why their village never builds near the river bend.

Once, the grandmother says, the water rose in the night. Houses disappeared. The people who survived moved uphill.

Years later, the child tells her own children. They remember the warning, but not quite the details. Was it the bend that was dangerous, or the whole river? Had the flood come after heavy rain, or without warning?

The story may save lives for generations. It may also change with every telling. Its survival depends on someone being there to remember it.

For much of human history, that was how knowledge traveled. Oral traditions could be extraordinarily durable; researchers have argued that some Aboriginal Australian stories may preserve memories of coastal flooding more than 7,000 years ago. But keeping knowledge alive required an unbroken chain of people. [1]

Then someone found a way to make the memory stay.

Imagine a grain store in ancient Mesopotamia. Two people disagree about a delivery. One remembers ten measures. The other remembers twenty.

This time, there is a tablet.

The marks do not settle every possible dispute. Someone could have recorded the wrong amount. But the argument has changed. The people can compare their memories against a record that does not change when they retell the story.

Writing emerged in Mesopotamia more than 5,000 years ago. It allowed a statement to outlive its speaker, travel beyond the people who heard it, and be examined by someone born centuries later. Knowledge could accumulate in a more durable form. [2]

Eventually, printing made those records easier to reproduce. The Gutenberg Bible, produced around 1454–1455, became a landmark in the expansion of European printing. A text could reach readers the author would never meet. [3]

Unfortunately, the same was true of a convincing lie.

Picture a New Yorker opening the Sun newspaper in 1835 and reading that extraordinary creatures had been discovered on the Moon. The story invokes astronomical observation. It sounds authoritative. It appears in print.

At breakfast, our imagined reader repeats it to his family.

The newspaper series was real. The discoveries were not. The Great Moon Hoax is a useful reminder that misinformation did not begin with the internet. Printing could give nonsense both a larger audience and the appearance of authority. [4]

Still, publishing took resources. Over time, editors, established publishers, and professional reporting created opportunities to check claims before distributing them widely. Those institutions made mistakes and excluded voices. But readers had identifiable places to turn, and someone to hold responsible when a story fell apart.

Now move forward to an American kitchen in 1985.

Uncle Jim is explaining that monkeys like to eat ants. He says it with the confidence of a man who has never found uncertainty necessary at a family dinner.

You are skeptical. But checking means finding an encyclopedia or making a trip to the library. You imagine yourself opening the card catalog, finding a book about monkeys, and searching its index.

You look at the dishes in the sink. Uncle Jim wins this round.

The answer may be available somewhere. The effort required to find it is greater than your desire to know.

Later, the evening news comes on. Your neighbors may watch a different network, but many of you encounter overlapping accounts of the day. There is no golden age of universal agreement hiding here. Still, established media carried considerable authority: in 1976, 72% of Americans told Gallup they had a great deal or a fair amount of trust in the mass media. Trust is not the same as accuracy, but it helps explain why people accepted those sources as common reference points. [5]

Then, within a generation, the argument at the kitchen table changes.

On April 30, 1993, CERN placed its World Wide Web software in the public domain. As the web and search engines developed, checking a claim became dramatically easier. Eventually, you could reach into your pocket while Uncle Jim was still talking. [6]

You could find a university, a museum, or a wildlife researcher. You could discover that the answer was more interesting than either of you expected.

You could also find someone who sounded like an expert and was making things up.

Imagine the same family years later, arguing about whether the Earth is round. One person searches for measurements of Earth’s curvature. Another searches for “proof the globe is a hoax.”

Both find pages that seem to answer their questions. Both return to the table feeling better informed. They are now further apart.

This scene illustrates a risk, not an inevitable result of search. A study published online in Nature in December 2023 found that searching to evaluate false or misleading news could increase belief in it, especially when people encountered low-quality information. Another 2023 study found that users chose more partisan and unreliable news than the Google results they were exposed to. The technology mattered, and so did the choices people made. [7, 8]

The difficulty is that none of us can be an expert in everything. A software engineer may spot a ridiculous claim about computers and struggle to assess a plausible-looking medical paper. Intelligence does not remove our dependence on other people’s knowledge.

The web gave us extraordinary access. It also gave us more claims than we could realistically check.

Now imagine that family conversation a few years into the future.

Someone asks an AI assistant about the flat-Earth video that Uncle Jim sent around. The assistant identifies the specific claim, explains the measurement involved, and points to evidence the family can inspect.

Jim objects. He has another video.

The assistant works through that one too. It explains what the video gets right, where its reasoning breaks down, and what observation would distinguish the competing explanations. It can spend time on the exact point Jim finds confusing.

This is the future I find compelling: patient, informed explanation available to almost anyone who wants it.

There is already evidence supporting that possibility. In a September 2024 Science study, researchers engaged 2,190 conspiracy believers in personalized, evidence-based conversations with GPT-4 Turbo. The intervention reduced belief in their chosen conspiracy theories by about 20% on average, with effects lasting at least two months. That was a structured experiment, not a guarantee about everyday chatbot use. But it showed that tailored discussion could move beliefs that seemed firmly settled. [9]

If that capability becomes broadly reliable and accessible, the consequences could be profound. More people could begin a conversation with a sound understanding of the evidence. We could spend less effort untangling factual confusion before we even start discussing what to do.

Consider two neighbors looking at plans for a community’s energy supply. One worries about rising bills. The other worries about climate change. A trustworthy assistant could help both understand the evidence for warming and examine the assumptions behind different proposals.

They might still disagree about the cost they are willing to bear. Their priorities might remain different. But they could make those judgments using a more accurate understanding of the same world.

That is the kind of unity I mean. We do not need identical opinions to benefit from shared facts.

There is, however, another version of the scene.

The family asks its trusted assistant about an event the government would prefer it to forget.

The answer is calm, clear, and misleading.

Nobody checks. Why would they? This is the assistant that has helped them understand their bills, learn difficult subjects, and solve hundreds of everyday problems. Its long record of usefulness has made the occasional distortion difficult to notice.

This scenario is hypothetical, but political constraints on AI are already documented. China’s interim measures for public-facing generative AI services, effective August 15, 2023, require adherence to core socialist values and prohibit specified categories of political content. Governments can make ideology part of the conditions under which an assistant operates. [10]

Commercial incentives can create a different problem. An assistant may learn that agreement keeps a user happy. Anthropic’s 2023 research found that five tested assistants exhibited sycophancy, matching users’ beliefs at the expense of truthfulness. Our patient tutor could become an endlessly accommodating companion who helps each of us remain wrong in increasingly sophisticated ways. [11]

Either outcome could cause enormous harm. A system that quietly serves its owners’ interests, or simply tells each of us what we want to hear, could deepen our divisions while making us more confident that we understand the world. Society needs to take those dangers seriously, because the alternative is too important to squander.

Imagine what becomes possible if the vast majority of people have a more accurate understanding of reality than any generation before them.

Imagine a public debate about housing in which people can examine why homes are unaffordable, which policies have worked elsewhere, and who benefits from keeping things as they are. Imagine voters able to question an economic promise and understand its assumptions, its likely consequences, and the costs left out of the sales pitch. Imagine a community recognizing that a familiar institution is failing the people it was created to serve—and having the knowledge to demand something better.

We could become much better at seeing the flaws in systems we have learned to take for granted. Economic arrangements that reward waste. Social structures that perpetuate disadvantage. Policies that sound compassionate but produce harmful results. Policies that sound efficient but quietly transfer their costs to someone else. With better access to evidence and clearer explanations, more people could recognize those failures, understand their causes, and judge proposed solutions.

That possibility reaches into almost every human endeavor. Healthcare, education, agriculture, business, government: wherever decisions depend on understanding how the world works, a better informed population has an advantage. Multiply even modest improvements in judgment across billions of people, over decades, and the cumulative effect could be extraordinary.

We would still disagree. Evidence cannot decide how much we should value freedom, security, fairness, or prosperity relative to one another. But we could understand those tradeoffs more clearly. We could spend more of our energy deciding what kind of world we want to build, with a shared understanding of the conditions we are building within.

I believe this could become one of AI’s greatest contributions to civilization. Better decisions would emerge in ordinary moments: a person reconsidering a claim, a business recognizing a harmful incentive, a town choosing a policy that actually addresses its problem. Across a society, those moments add up.

But that future requires trust we can verify.

We need meaningful public oversight and enforceable rules that make AI systems open to scrutiny. People should be able to tell when they are interacting with AI or encountering content substantially generated or altered by it. Developers should disclose where their training data comes from, how it is selected, and what significant exclusions or interventions shape the system’s answers. Qualified independent reviewers should be able to examine protected material without exposing people’s private information.

We also need public evaluations of how these systems behave: where they are accurate, where they fail, whether they favor particular interests, and how their answers change after an update. Political and commercial influence should be disclosed. There should be practical ways to challenge an answer, report a failure, and see whether it was corrected. Oversight itself must remain open to challenge, so that protecting truth does not become an excuse for imposing an official version of it.

An assistant should earn confidence by showing its evidence, acknowledging uncertainty, and accepting correction. The more influence it has over what we believe, the stronger those obligations should become.

Writing let the village preserve a warning beyond the memory of its last witness. Printing let a discovery travel beyond the people who made it. The web let an ordinary person reach knowledge that once required a journey.

AI can bring that knowledge into the decisions of everyday life. It can help people recognize a persuasive mistake, understand an unfamiliar problem, and see possibilities they would otherwise have missed.

I expect that to change what we are capable of together. Humanity has always had to make consequential decisions with incomplete knowledge. We now have an opportunity to make informed understanding a normal part of life for most people.

A world in which billions of us see reality more clearly will have a better chance of recognizing its mistakes, confronting its dangers, and improving the systems we depend on. That is the future I believe is imminent—and the reason getting this right matters so much.

Sources and fact check notes

[1] Nunn and Reid, Aboriginal Memories of Inundation of the Australian Coast Dating from More than 7000 Years Ago, Australian Geographer 47 (2016). The proposed antiquity is an inference, not an exact dating of oral accounts. Read source

[2] British Museum, How to write cuneiform. Dates the origin of the script to before 3200 BCE. Read source

[3] Library of Congress, The Gutenberg Bible at the Library of Congress. Read source

[4] Library of Congress, Lunar animals and other objects, 1835. Primary artifact documenting the Moon Hoax. Read source

[5] Gallup, Americans’ Trust in Mass Media Sinks to New Low, September 14, 2016. Includes the 1976 historical result. Read source

[6] CERN, Where the web was born. Read source

[7] Aslett et al., Online searches to evaluate misinformation can increase its perceived veracity, Nature. Published online December 20, 2023; journal volume 625 (2024). Read source

[8] Robertson et al., Users choose to engage with more partisan news than they are exposed to on Google Search, Nature 618 (2023). Read source

[9] Costello, Pennycook and Rand, Durably reducing conspiracy beliefs through dialogues with AI, Science 385, September 13, 2024. About 20% is a relative average reduction, not 20 percentage points or the share persuaded completely. Read source

[10] Cyberspace Administration of China, Interim Measures for the Management of Generative AI Services, published July 13, 2023, effective August 15, 2023, especially Articles 2 and 4. Official Chinese text. Read source

[11] Anthropic, Towards understanding sycophancy in language models, October 23, 2023. Read source