Skip to content
OGERIA — Observatory of Global Evidence on Risks in AISynthesis report · 2026 ed.
Updated 2 Oct 2026

Chapter 02 · Vectors

Epistemic and information

Degradation of the collective ability to tell true from false, and of the trust that sustained it.

This vectorVectorThe path by which a risk moves from the screen into the world: biological, cyber, military, economic, political, epistemic or loss of control. In this observatory, each vector has its own colour.For exampleA burglar can get in through the door, the window or the roof. The burglar is the risk; the door, the window and the roof are the vectors. is defined by what it is not. The dominant fear —mass personalised persuasion swinging elections— is the worst-supported part. In preregisteredPreregistered studyA study that published, before starting, what it would measure and how it would analyse it, so that it cannot later pick only the results that suit it.For exampleLike announcing which number you are betting on before rolling the dice, not after seeing how they landed. experiments, GPT-4’s personalisation wins the debate 64.4% of the time —an 81.2% increase in relative odds, not 81% more persuasion— and the correction the journal itself published in September 2026 also struck down the part most often cited: the advantage of personalising over not personalising stops being significant (P = 0.07) [931]Author Correction: On the conversational persuasiveness of GPT-4Salvi, Francesco; Horta Ribeiro, Manoel; Gallotti, Riccardo et al. · 2026 · paperView source ↗Accessed on 9 September 2026. In the largest study available post-training is the lever that raises persuasiveness the most, at the cost of factual accuracy: “where they increased AI persuasiveness, they also systematically decreased factual accuracy” [493]The levers of political persuasion with conversational artificial intelligenceHackenburg, Kobi; Tappin, Ben M.; Hewitt, Luke et al. · 2025 · paperView source ↗verified through CrossrefAccessed on 9 September 2026. In the census of 2024 electoral incidents, “cheap fakes” were used seven times more than AI-generated content [585]We Looked at 78 Election Deepfakes. Political Misinformation Is Not an AI Problem.Kapoor, Sayash; Narayanan, Arvind · 2024 · institutional blogView source ↗Accessed on 9 September 2026.

What is documented enters from another side, and it is direct harm to people: between the end of December 2025 and 8 January 2026, two independent measurements estimate between 1.8 and 3 million sexualised images generated of real people, including minors [112]Complaint — Mayor and City Council of Baltimore v. X Corp., x.AI Corp., x.AI LLC and Space Exploration Technologies Corp. (Case C-24-CV-26-002129)Thompson, Ebony M.; Baltimore City Law Department · 2026 · official documentView source ↗Accessed on 9 September 2026. The range between the two figures is information in itself: nobody has an exact count.

The underlying mechanism is not deception either, but the erosion of the guarantee: when any record could be false, whoever lies can dismiss what is true. The effect exists and is bounded —claiming disinformation works after text-based scandals and not against video [939]The Liar's Dividend: Can Politicians Claim Misinformation to Evade Accountability?Jackson Schiff, Kaylyn; Schiff, Daniel S.; Bueno, Natália S. · 2025 · paperView source ↗Accessed on 9 September 2026. Synthetic volume, meanwhile, has already brushed half the measurable open corpus: 50.9% of articles in the fourth quarter of 2025 and 49.9% in the first quarter of 2026 [472]AI Now Writes as Many Online Articles as Humans DoGraphite · 2026 · reportView source ↗Accessed on 9 September 2026.

Risks in this vector

  • SpeculativeEpistemic dependenceIrreversible

    Delegating belief formation to a system that affirms more than it corrects, until the ability to discern for oneself atrophies.

  • ProjectedContamination of the public recordSevere

    Synthetic material enters the shared record -web, encyclopedias, literature- and stops being distinguishable from the original.

  • ObservedAutomated electoral disinformationSevere

    Producing false political content stops costing money and time, right in the largest electoral cycle in history.

  • ProjectedCultural homogenisationSevere

    If a growing share of what is read, seen and heard passes through a few models, the available repertoire narrows.

  • LabPersonalised persuasion at scaleSevere

    Conversations optimised to convince, held with millions of people at once, without it showing that a model is on the other side.

See among the vectors →Report a mistake in this entry →

Sources

  1. [931] Author Correction: On the conversational persuasiveness of GPT-4 · Nature Human Behaviour 2026
  2. [493] The levers of political persuasion with conversational artificial intelligence · UK AI Security Institute / University of Oxford 2025 verified through Crossref
  3. [585] We Looked at 78 Election Deepfakes. Political Misinformation Is Not an AI Problem. · Knight First Amendment Institute, Columbia University 2024
  4. [112] Complaint — Mayor and City Council of Baltimore v. X Corp., x.AI Corp., x.AI LLC and Space Exploration Technologies Corp. (Case C-24-CV-26-002129) · Circuit Court for Baltimore City 2026
  5. [939] The Liar's Dividend: Can Politicians Claim Misinformation to Evade Accountability? · Purdue University / Emory University 2025
  6. [472] AI Now Writes as Many Online Articles as Humans Do · Graphite 2026

Ask OGERIA

It answers only with what the observatory publishes and can be wrong: check the entries it cites. Your questions are sent to an AI model, so don't write personal data. More in the privacy policy.

Up to 500 characters.

Support OGERIA on Ko-fi

The payment is processed by Ko-fi, not by this site. Open on ko-fi.com