Engagement algorithms favor misinformation

The original insight

Don’t let outliers skew your opinion and actions.

What “manipulation is the game” means

In this formulation, manipulation means the deliberate shaping of human behavior, not necessarily deception or coercion.

Engagement systems are designed to influence what people do next. They select what appears in front of each person, observe what captures attention, test which subjects and presentations produce a response, and continually adapt the feed to increase watching, clicking, reacting, sharing, returning, and continued use.

Influencing the user’s next action is not an accidental side effect of an engagement algorithm. It is a central function of the system.

That does not mean the algorithm has a mind, a motive, or a conscious desire to mislead. The intention exists at the level of the system’s design: optimize the environment so that it changes behavior in the direction the platform rewards.

The manipulation is intentional at the design level. The amplification of misinformation is an emergent result.

Why misinformation is the pain

Engagement algorithms favor misinformation indirectly. They do not need to recognize it, seek it, or be instructed to promote it. They simply optimize for the behavior it can produce.

When a video, article, or social-media post reaches us, it is easy to behave as though it appeared because it was important or well supported. Usually it appeared because an algorithm predicted that we would watch, click, react, or continue scrolling.

Truth and attention are not opposites. Many accurate videos are interesting, and platforms apply separate policies and authority signals in some sensitive areas. But evidentiary reliability is not the primary universal measurement presented to the viewer.

Falsehood has an attention advantage

A video confirming what the wider evidence already shows has little novelty. A video claiming that experts, institutions, or common understanding are completely wrong has a ready-made hook. Contradiction stands out because it is unusual.

An outlier can be completely true and still create a false impression. A single video, personal testimony, or extraordinary case can outweigh thousands of less dramatic observations in the viewer’s mind. The story need not be false to mislead; it need only be unrepresentative.

When a video makes an accepted belief suddenly appear foolish, corrupt, or completely mistaken, pause before absorbing its conclusion. Ask whether it has uncovered strong contradictory evidence or merely packaged the most persuasive exception.

AI makes verification practical

Historically, checking every influential claim properly required time, research skill, access to sources, and more patience than most people could reasonably devote to every video they encountered. AI changes that practical constraint.

A capable AI with web access can extract important factual claims, search for relevant evidence, compare them with the wider record, identify missing context, distinguish fact from opinion or prediction, and present the sources within minutes.

A useful truth or verifiability score need not declare that an entire video is simply true or false. It could measure whether the central claims are identifiable, verifiable, supported by reliable evidence, representative of the wider data, and presented with necessary context.

Imperfection would not make the score meaningless. It could be tested, audited, challenged, and improved. Its purpose would be to create evidentiary friction before belief, not to claim omniscience.

Protecting legitimate novelty

A radical new idea should not be labelled false merely because it contradicts current consensus. “Unverified,” “emerging,” “supported by limited evidence,” and “contradicted by strong current evidence” are different assessments.

The score should expose those distinctions rather than enforce established dogma. It should label evidence, not suppress ideas.

Why isn’t there a “How true is this?” button?

Why does every YouTube video not include a prominent How true is this? button or an expandable evidence label?

Google already operates Fact Check Explorer. YouTube already supplies contextual information panels for selected topics and says it uses human evaluation and machine-learning systems to reduce some borderline misinformation. The technical and institutional pieces therefore exist in partial form.

The most interesting obstacle may be the incentive. Imagine that many highly compelling videos acquired visible badges showing low verifiability, weak evidentiary support, or serious missing context. Would viewers become better informed and trust the platform more over time? Or would they click less, abandon sensational videos sooner, and reduce total viewing?

No public evidence reviewed here establishes that Google deliberately withheld such a feature to protect views. But the incentive question remains legitimate: when a system is rewarded for capturing attention, how much commercial motivation exists to place immediate evidentiary friction beside every compelling claim?

Falsehood is not the objective

No executive needs to prefer misinformation. No conspiracy is required. If the strongest rewards concern engagement while truth is harder to measure and may interrupt consumption, the system will naturally become better at predicting what people will watch than at helping them decide what they should believe.

Claims that contradict established reality often possess exactly the qualities that attract attention: they are novel, surprising, alarming, and easily framed as revelations that overturn everything we thought we knew.

The algorithm does not need to recognize a claim as false. It only needs to observe that people stop scrolling, click, watch, react, and share. It then rewards the behavior generated by the radical claim.

The untruth is not the goal. It is the content most naturally fitted to exploit the goal.

The corporate-incentive analogy

Food companies do not need to set out to cause type 2 diabetes. They pursue growth, repeat purchasing, market share, and profit. Harmful population-level consequences can emerge from the products and behavior those incentives favor.

In the same way, a recommendation algorithm does not need to set out to spread misinformation. It pursues attention, continued viewing, and engagement. Misinformation can emerge disproportionately because a falsehood presented as a hidden truth can be more compelling than the familiar reality it contradicts.

In both cases, harmful intent is unnecessary. The damage is an emergent property of the incentive structure.

Blaming the user is a cop-out

The companies behind these systems can offer a technically plausible defense: the algorithm did not select the video because it contained misinformation; it responded to what users chose to watch.

But this isolates the final click from everything that shaped it. The platform selected which videos entered the menu, where each appeared, how often it reappeared, what followed automatically, which thumbnail and headline accompanied it, and which signals determined its further distribution.

The same defense appears in the food system. Nobody forces a person to buy junk food. Yet companies design it for repeated consumption, surround people with advertising, give it prominent placement, make it convenient, and then point to the final purchase as though it occurred in a neutral marketplace.

Personal agency remains real. But blaming the viewer for clicking or the consumer for buying is a cop-out when the company constructed the choice environment, optimized it to influence behavior, and profits from the predictable result.

Which button would you press?

Imagine being offered two buttons. The first provides material that remains compelling and entertaining but is also oriented toward reality. The second provides whatever is most likely to capture and hold your attention, with little concern for whether it leaves you accurately informed or completely misled.

Entertain me, but help me understand what is real.

Or:

Entertain me; truth is optional.

Most people would presumably choose entertainment that also keeps them oriented toward reality. Yet current platforms rarely offer that choice explicitly. They present the engagement-optimized stream first and leave verification as extra work for the viewer.

And if you knowingly choose the second button, you might also be interested in a truly exceptional deal on a bridge in London.

A practical AI fact-checking prompt

Audit the factual claims made in this video. First list every material claim and classify it as factual, opinion, prediction, interpretation, or rhetoric. For each factual claim, search current reliable sources, prioritizing primary evidence, and give direct links. Rate the claim as supported, partly supported, unsupported, contradicted, or not verifiable, and explain why. Identify missing context, cherry-picked examples, base-rate neglect, and evidence that points in the opposite direction. Distinguish whether an example is true from whether it is representative. State important uncertainty and do not invent a verdict where the evidence is insufficient.