They Sounded Crazy - Until the Internet Proved Them Right. What That Reveals About AI Today

History offers an attractive shortcut when a technology unsettles us. Find someone who underestimated the internet, place their quotation beside a confident prediction about AI, and the conclusion seems to follow: doubters will be wrong again. Reverse the selection and the same history can support a warning that every new technology will produce unexpected harm.

Neither argument is sufficient. We should learn from the internet without pretending that its history predicts AI’s trajectory. The useful lesson concerns how institutions respond to technologies whose applications, incentives and consequences are still changing.

A failed prediction is not a forecasting method

People made conflicting predictions about the internet. Selecting the most memorable mistakes after the fact tells us little about the reliability of a new forecast. We also need to ask which predictions were accurate, which were too vague to test and which depended on conditions that later changed.

The original World Wide Web project page is a useful historical artefact. It describes a linked information system and invites participation. It documents a project’s aims; it does not establish that its creators either foresaw every consequence or understood none of them.¹

That distinction protects the argument from an easy mistake. A developer can understand an immediate technical purpose while remaining uncertain about future institutions and business models. The resulting uncertainty does not make every later outcome unforeseeable. It means that foresight needs to be updated as evidence and incentives become clearer.

For AI, I would therefore separate three kinds of statement: a capability demonstrated under specified conditions, a forecast about later capability, and a claim about social consequences. A persuasive account should explain the steps between them. A striking demonstration cannot, on its own, tell us which occupations will change, which harms will grow or which policies will succeed.

Expertise deserves attention, not exemption

Warnings from researchers and industry leaders matter because these people may understand particular mechanisms and limitations. Their forecasts still need examination. Expertise in building a model does not automatically confer a precise probability for a future social outcome.

The 2023 statement on AI extinction risk documents a public position: its signatories argued that mitigating that risk should be a global priority. It is not a measurement of the probability of extinction, nor a demonstration that all relevant experts agree on a timeline or mechanism.²

We should neither dismiss a warning because it is difficult to quantify nor treat a prominent signature as proof. Ask what could produce the outcome, which assumptions it requires, what evidence bears on those assumptions and what intervention might reduce the risk. The same discipline should apply to confident reassurances.

Motives should also be handled carefully. An industry participant can possess important knowledge and have commercial interests at the same time. An observer’s suspicion of those interests does not establish that a warning is insincere; the participant’s technical experience does not eliminate the need for independent assessment. Evidence and accountability are more useful than guessing who is secretly afraid or selling a story.

Present harms and uncertain futures need different responses

The International AI Safety Report 2026 describes uneven capabilities, documented harms and substantial uncertainty about future development. Its discussion distinguishes current evidence from scenarios involving more advanced systems. That is a more useful starting point than declaring either that everything is already known or that nothing can be done until the future arrives.³

There are different questions to investigate. Does a deployed service produce unreliable information in a consequential workflow? Can tools and permissions turn a model error into an external action? Could later systems create hazards that present evaluations do not capture? Those questions should not be merged into one undifferentiated claim that AI is safe or dangerous.

The distinction is practical. A known deployment failure may call for correction, independent investigation or a change in use. An uncertain future capability may call for research, scenario exercises and evaluation triggers. A sweeping ban and an unconditional rollout are not the only available choices.

Employment deserves the same care. A fall in recruitment in a particular period cannot be assigned entirely to AI without considering the economy, industry conditions and employer choices. An early signal in one occupation should not become a confident forecast for all work. Workers nevertheless have good reason to demand information about changes affecting them before those changes become irreversible.

Who participates in the decision?

Public participation is often invoked after an organisation has already chosen its objective. That is too late if the choices determine who receives a benefit, who is monitored or who can contest an outcome. A consultation should state which decisions remain open and how contributions can change them.

My proposed approach would bring affected groups into discussions about the intended use, acceptable failure and available remedies. Technical testing remains necessary. It answers a different question from whether a community considers a use legitimate. Neither expertise nor participation should be used as a ceremonial substitute for the other.

Oversight also needs authority. A reviewer who can document a concern but cannot pause deployment has limited influence. Before adopting a consequential system, an institution should name the people able to restrict, reverse or discontinue its use, and the evidence that would trigger such action.

This is where the internet analogy becomes useful. Instead of claiming that history guarantees a particular outcome, ask how decisions became difficult to reverse, which business incentives mattered and when affected people obtained meaningful alternatives. Those are research questions to investigate in each context, rather than a prewritten verdict on AI.

Build an account that can change

I would require a public-facing decision record for consequential deployments: the intended use, the evidence considered, the unresolved uncertainties, the responsible institution and a date for reassessment. The document should separate results from promises and show what changed between reviews.

Such a record need not disclose sensitive operational details or personal information. It should make the public claim inspectable. If an institution says a system improves fairness, readers should be able to identify the comparison and its limits. If it says that human review protects users, it should explain what that review can actually do.

The aim is not to demand perfect prediction. It is to create conditions under which mistakes can be recognised and corrected. A policy that can respond only after a catastrophic failure is weak. So is a policy built on a forecast that cannot be questioned.

What history can responsibly offer

The internet’s history teaches humility about prediction and attention to institutions. It does not establish that AI will repeat the same path, that the most ambitious forecasts will prove correct or that its worst scenarios are inevitable.

We can take warnings seriously while asking for evidence. We can recognise beneficial capabilities while questioning the terms of deployment. The strongest position is one that can explain what it knows, what it does not know and how its judgment will change when the evidence changes.

That is a better inheritance from technological history than another collection of people who sounded wrong before events made them famous.

Sources and notes

1. CERN, original World Wide Web project page. Used as a historical artefact of the project’s stated purpose, not a complete account of its creators’ predictions. ↩
2. Statement on AI Risk (May 2023), signatory statement. A documented policy position, not a probability estimate or a universal scientific consensus. ↩
3. International AI Safety Report 2026, Executive Summary (February 2026). Evidence synthesis distinguishing documented harms, capability limits and uncertain future scenarios. ↩