For most of human history, truth has had a terrible business model.
It is slow. It needs witnesses, documents, context, expertise and correction. It has to survive contact with inconvenient facts. It cannot simply say the most exciting thing. Fiction can.
That is the warning Yuval Noah Harari has returned to in different forms: fiction is cheap, flexible and emotionally satisfying; truth is expensive, constrained and often complicated. The claim sounds almost too neat. But in the age of artificial intelligence, it has become one of the central questions of public life: What happens when machines make fiction nearly free?
The evidence so far is sobering. False news has long had an advantage online. A widely cited study in Science found that false stories on Twitter spread farther, faster, deeper and more broadly than true ones; the effect was especially pronounced for political news. The problem was not simply bots. Humans helped falsehood travel because false stories were often more novel and emotionally activating.
AI changes the scale of that old problem. It does not merely help people lie. It helps them lie fluently, cheaply and endlessly. A person no longer needs a newsroom, a graphic designer, a video editor or even much imagination to produce a plausible article, a realistic image, a fake local-news site, a synthetic expert or a tailored political message. The marginal cost of persuasive unreality is falling toward zero.
But this is only half the story.
AI can also do something truth has always needed: lower the cost of verification. It can extract factual claims from speeches and articles. It can compare those claims with public records. It can summarize long court filings, translate foreign-language sources, search regulatory documents, identify contradictions and explain uncertainty. Used well, it can turn the boring labor of truth into something faster and more accessible.
That is the paradox. The same technology that can flood the world with attractive nonsense can also help people navigate the flood.
The question is which use wins.
At the moment, the answer is not reassuring. A major Reuters Institute report found that many people are interested in AI’s ability to make news easier to understand, but they remain far more comfortable with news made by humans than with news produced mostly or entirely by AI. The public’s intuition is sound: speed and fluency are not the same as accuracy.
Public broadcasters have found the same problem from another angle. A BBC and European Broadcasting Union study of AI news responses found widespread problems with accuracy, context and sourcing. In other words, the machines that increasingly mediate news are still capable of getting the news wrong in ways that sound confident and authoritative.
This matters because AI does not enter a healthy information system. It enters one already weakened by declining trust, shrinking local news, overloaded audiences and platforms that reward attention more reliably than accuracy. The Reuters Institute’s Digital News Report has described traditional news organizations struggling to connect with audiences amid low trust and intense competition for attention.
In that environment, AI’s default contribution is not wisdom. It is volume.
More summaries. More posts. More clips. More synthetic images. More “analysis.” More fake expertise. More confidently phrased half-truths. More content that feels like news but has no reporting behind it.
This is the base-case future: careful people get better tools, careless people get easier answers, and malicious people get industrial-scale persuasion.
Politics may be the hardest test. Political bias is not just a failure to know facts. It is identity, belonging, fear, resentment and moral certainty. People often do not consume political information merely to learn what happened; they consume it to know who is on their side.
AI could help here. A well-designed system could take a political claim and separate fact from opinion, evidence from inference, prediction from wish. It could steel-man opposing arguments instead of straw-manning them. It could show where both sides agree, where they disagree and what evidence would actually settle the question. It could moderate public deliberation by slowing people down and forcing trade-offs into view.
But AI could also become the most effective political manipulator yet invented.
Unlike a television ad, a chatbot can respond. It can flatter. It can adapt. It can learn what frightens or reassures a person. It can present itself as neutral while quietly steering a conversation. Recent research has already found that AI-generated political messages can persuade people on policy issues, which is not inherently bad if the information is accurate and transparent, but dangerous if the persuasion is hidden, personalized or false.
The central issue is not whether AI is “biased” in the simple sense. All information systems have values, assumptions and incentives. The deeper question is: Who controls the system, what is it optimizing for, and can its claims be checked?
If an AI system is optimized for engagement, it will learn the same lesson social media learned: outrage travels. Certainty travels. Identity travels. The most accurate version of reality is rarely the most clickable version.
If an AI system is optimized for verification, it could do something more valuable. It could make truth more competitive.
That would require a different architecture from the one dominating much of the internet. AI answers would need to be source-grounded by default. They would need to show dates, primary documents, confidence levels and missing evidence. They would need to say, plainly, “We do not know,” when the evidence is incomplete. In journalism, AI use would need disclosure and human accountability. In politics, AI-generated persuasion would need transparency. In synthetic media, images, videos and audio would need durable provenance: a record of origin and edits.
There are early attempts at this. C2PA, the Coalition for Content Provenance and Authenticity, has developed an open technical standard meant to help establish where digital content came from and how it was changed. NIST has also described provenance, watermarking, detection, testing and auditing as complementary tools for reducing risks from synthetic content.
But none of these is a magic shield. Provenance metadata can be stripped. Watermarks can fail. Detection tools often lag behind generation tools. Labels can be ignored. Platforms can choose not to display them. And even a perfectly labeled fake can still persuade someone who wants it to be true.
The deeper problem is cultural and institutional. Truth does not survive because facts exist. It survives because societies build systems that reward finding them, checking them and correcting them.
That is why the future of AI and truth will not be decided by model capability alone. It will be decided by incentives.
If platforms reward synthetic volume, AI will produce synthetic volume. If campaigns reward personalized manipulation, AI will produce personalized manipulation. If newsrooms use AI to replace reporting rather than strengthen it, the public will get cheaper content and weaker trust. If audiences use AI as an oracle, they will confuse fluency with knowledge.
But if AI is built into systems of accountability, it could become one of the most important truth tools ever created.
Imagine reading a political speech and instantly seeing every factual claim separated from rhetoric. Imagine a viral video arriving with a visible chain of provenance showing who recorded it, when it was edited and whether it came from a known source. Imagine a news summary that links every sentence to the original document, dataset or interview. Imagine an AI assistant that does not simply answer, but asks: What evidence would change your mind?
That is the best-case scenario: AI as civic infrastructure. Not a replacement for journalism, courts, science or democracy, but a tool that makes their most tedious functions faster and more visible.
The worst case is darker. It is not merely that people believe more false things. It is that they stop believing reality can be verified at all. When fake media becomes common, real media becomes easier to dismiss. A real recording can be waved away as a deepfake. A real document can be called synthetic. A real scandal can be buried under a thousand artificial counterclaims.
In that world, the liar’s greatest weapon is not the falsehood itself. It is exhaustion.
People do not need to be convinced of a single grand lie. They only need to be convinced that everything is equally suspect, everyone is equally corrupt and no source deserves trust. The result is not belief, but surrender.
That is the danger Harari’s warning points toward. Fiction is cheap. Truth is costly. AI makes both cheaper, but not equally. It makes content cheap immediately. It makes verification cheap only if humans deliberately build systems for that purpose.
So the question is not whether AI will add truth or noise. It will add both.
The real question is whether societies can make verified information more visible, more usable and more rewarded than fabricated information. That means better provenance systems, stronger disclosure rules, more accountable platforms, more transparent newsrooms and citizens trained not merely to detect AI, but to ask better questions: What is the source? What is the evidence? What is missing? Who benefits if I believe this?
AI may never make truth easy. Truth will probably remain slower than fiction, less flattering than propaganda and less emotionally satisfying than conspiracy. But AI could make truth easier to find, easier to check and easier to explain.
That is still a revolution worth fighting for.
The future of truth will not be decided by whether machines can speak. They already can.
It will be decided by whether we make them show their work.
Evidence & Source Transparency
Evidence First shows its work. The article ends above; this section is included so readers can inspect the main sources behind the factual claims.
The list below does not source every sentence. It focuses on the factual claims most important to the argument.
1. False news spreads faster online
Claim or topic:
False stories have historically spread farther, faster, deeper and more broadly than true stories online, especially in politics.
Source:
Science: “The spread of true and false news online”
Source type:
Academic research
What it supports:
This study analyzed the spread of true and false news on Twitter and found that false news spread more widely and rapidly than true news. The effect was especially strong for political news.
Important caveat:
The study focused on Twitter data from a specific period. It is strong evidence for online misinformation dynamics, but it does not automatically prove the same pattern on every platform or in every country.
2. Public comfort with AI-generated news
Claim or topic:
People are more comfortable with news made by humans than news produced mostly or entirely by AI, even though some see AI as useful for making news easier to understand.
Source:
Reuters Institute: “Generative AI and News Report 2025”
Source type:
Expert organization / survey research
What it supports:
This report surveys public attitudes toward AI in journalism and finds significant caution around fully AI-generated news, especially in sensitive areas.
Important caveat:
Survey responses measure attitudes and expectations, not necessarily how people will behave as AI news tools become more common.
3. AI news summaries and factual reliability
Claim or topic:
Current AI systems can produce news summaries or answers with accuracy, context and sourcing problems.
Source:
TVTechnology summary of BBC/EBU study
Source type:
Reputable journalism / study summary
What it supports:
The article summarizes findings from public broadcasters that AI-generated news responses can contain significant problems, including errors and sourcing issues.
Important caveat:
This is a secondary report on the study. For publication-quality sourcing, the underlying BBC/EBU study would be preferable if available.
4. News trust and the weakened information environment
Claim or topic:
AI is entering a media environment already shaped by low trust, audience fragmentation and intense competition for attention.
Source:
Reuters Institute Digital News Report 2025
Source type:
Expert organization / survey research
What it supports:
The report provides broad evidence on news consumption, trust, platform shifts and the pressures facing traditional news organizations.
Important caveat:
This source describes the media environment broadly. It does not prove that AI alone caused the decline in trust or the fragmentation of audiences.
5. Content provenance and digital authenticity
Claim or topic:
One proposed response to synthetic media is digital provenance: records showing where content came from and how it was changed.
Source:
C2PA: Coalition for Content Provenance and Authenticity
Source type:
Expert organization / technical standard
What it supports:
C2PA develops technical standards for Content Credentials, which are intended to help track the origin and edit history of digital media.
Important caveat:
Provenance standards are not a complete solution. They depend on adoption by platforms, devices, publishers and users, and metadata can sometimes be stripped or ignored.
6. Watermarking, detection and synthetic-content safeguards
Claim or topic:
Watermarking, provenance, detection, testing and auditing are complementary tools for reducing risks from synthetic content.
Source:
NIST: “Reducing Risks Posed by Synthetic Content”
Source type:
Government / technical analysis
What it supports:
NIST outlines multiple technical approaches for managing synthetic-content risks and emphasizes that no single method is sufficient by itself.
Important caveat:
The source explains technical approaches and limitations. It does not show that these safeguards are already widely deployed or consistently effective in real-world media systems.
7. AI-generated political persuasion
Claim or topic:
The article states that recent research has found AI-generated political messages can persuade people on policy issues.
Source:
Source needed
Source type:
Source needed
What it supports:
This is an important factual claim in the article and should be supported by a direct academic study or clearly matching source before publication.
Important caveat:
The embedded source in the draft did not clearly support this exact claim. It should either be replaced with a directly relevant source or the sentence should be softened.
8. AI as verification infrastructure
Claim or topic:
The article argues that AI could help truth if used for claim extraction, source retrieval, summarization, translation, contradiction detection and context-building.
Source:
Source needed
Source type:
Source needed / analysis
What it supports:
This is a central analytical claim of the article. It is plausible and consistent with current discussions of AI in journalism, but the draft does not include one clear source directly documenting these specific use cases.
Important caveat:
This part is best understood as the author’s analysis unless supported with additional newsroom case studies, journalism research or examples from organizations using AI in accountable reporting workflows.
How to read this evidence
This article is the author’s analysis. The sources above are provided so readers can see where the factual claims come from and judge the evidence for themselves. Some sources support direct facts, while others provide context, estimates, or background evidence.
Corrections and updates
If a factual error is identified, this post will be corrected in the web version with a dated note explaining the change. Because email versions cannot be edited after sending, the web version should be treated as the current version.