The World Economic Forum’s Global Risks Report 2026 ranks Misinformation and disinformation the second most severe global risk over the next two years and fourth over ten, Societal polarization third, Cyber insecurity sixth, and Adverse outcomes of AI the fastest-rising risk of all, from #30 to #5 across the two horizons. Its diagnosis is that telling authentic content from synthetic “is becoming progressively more difficult” (p34) and may be impossible within a decade (p64).
This paper takes that diagnosis as its starting point. When the artefact proves nothing, what remains measurable is behaviour: who moves a narrative, how fast, how synchronised, through which infrastructure, and whether the movement is organic or manufactured. We map five of the report’s top risks to the behavioural signals we measure in production and give the evidence we hold for each, including a published investigation that found no coordination, with its controls. We state the bounds on every claim and what we will publish next.
1. What the Report ranks
| Risk | Likeliest crisis in 2026 | Two years (2028) | Ten years (2036) |
|---|---|---|---|
| Geoeconomic confrontation | #1 (18%) | #1 | #19 |
| Misinformation and disinformation | #5 (7%) | #2 | #4 |
| Societal polarization | #4 (7%) | #3 | #9 |
| Cyber insecurity | #9 (3%) | #6 | #8 |
| Adverse outcomes of AI technologies | #8 (4%) | #30 | #5 |
Three movements matter. Misinformation moved to #2 only because the trade war rose to #1; it did not fall. Adverse outcomes of AI “is the risk with the largest rise in ranking over time” (p9). And geoeconomic confrontation is the only risk the experts expect to ease by 2036 (p19). The economic weapons change; the information weapons stay.
The report does not treat these as separate. “One of the strongest interconnections in the GRPS is between Societal polarization and Misinformation and disinformation” (p33). Trust in news fell from 44 percent to 40 percent between 2018 and 2025 while news avoidance rose from 32 to 40 percent (p35, citing the Reuters Institute). In the United States, social media as the primary news source rose from 4 percent in 2015 to 34 percent in 2025 (p36).
2. The premise: when content stops being evidence, behaviour remains
Content classifiers answer “is this false” by pattern-matching against what past falsehoods looked like. They degrade against generated text and against operators who adapt to the features that caught them. Behaviour does not have that weakness. A campaign that needs a narrative to move must still move it, and movement leaves timing, synchrony, repetition, account structure and infrastructure traces that an operator cannot remove without ceasing to be a campaign.
That is the method AI Uniti has run in production since March 2026. PulseCheck scores an account on a continuous bot-to-human spectrum from nine deterministic behavioural signals, each named in the verdict. Signal collects across X, Bluesky, Mastodon, YouTube and RSS, and runs coordination scans whose thresholds come from the monitored surface’s own baseline, not a fixed number carried in from somewhere else. The report’s forecast that synthetic content becomes ubiquitous is not a threat to this method. It is the condition the method was built for.
The discipline in full: what is narrative threat intelligence, and the method behind it, what is behavioural threat intelligence.
3. Five risks, what we measure, what we can show
| WEF risk (2028 rank) | The mechanism the report describes | What behavioural detection measures | Our evidence (as at the date shown) |
|---|---|---|---|
| Misinformation and disinformation (#2) | False content “widely spread through media networks” (p76) | Whether the spread is coordinated, automated or organic, with the surface’s own null as the yardstick | A published null on Australian political amplification: 9.9% coordination density against 19.5% and 39.3% for controls, confirmed by network topology and burst timing (July 2026) |
| Societal polarization (#3) | Divisive narratives amplified by technology (p16, p33) | When a narrative is built relative to the event it precedes | Ceuta crossings, July 2026: narrative peaked about fourteen days before the physical surge; distribution infrastructure predated the trigger by at least 22 months |
| Adverse outcomes of AI (#30 rising to #5) | Deepfakes cheaper and more convincing; disinformation “at scale” (p36, p66) | Carriage of synthetic media as one signal among the behavioural ones; the behaviour of automated and agentic accounts | A financial-fraud operation presenting as 58% positive sentiment; organic sentiment about 22% positive once the coordinated layers were removed (April 2026) |
| Cyber insecurity (#6) | Attacks on infrastructure and business with plausible deniability (p16, p51) | The narrative layer that accompanies and precedes intrusions: the synthetic executive, the manufactured complaint wave | Behavioural bot filtering on our own outbound traffic separated human from automated clicks in production; we do not detect intrusions and do not claim to |
| Geoeconomic confrontation and trust (#1) | Trade, finance and technology “wielded as weapons of influence” (p15); reputational risk central to corporate risk (p38) | Narrative pressure on organisations whose standing is an economic asset | Production scale as at 15 October 2026: 503,226 posts, 151,931 author profiles, 19,962 author pairs tested, 90 brands, 109 coordination scans, 1,235 PulseCheck runs, across every monitor on the platform including our own research monitors |
The null we published
In mid-2026 public commentary asserted a coordinated operation behind a surge in Australian political amplification. We reconstructed about 189,000 X posts and 8,600 YouTube items across four research brands with controls and tested the claim with three independent methods. On coordination density the target cluster was the lowest of the active set. Its interaction network was the least dense, with zero mutual reply triads across 809 active accounts. Its burst timing was indistinguishable from its mainstream-party control. We published the result rather than the exposé that would have served our marketing better. The bound is stated with it: a null for these methods on X and YouTube, not proof of absence. A detection company that publishes “nothing found, with evidence” is demonstrating the property buyers need: verdicts that can be trusted in both directions.
The window we measured
Between 30 and 31 July 2026, 50,000 to 60,000 people crossed from Morocco into Ceuta on the strength of a distorted one-sentence rendering of a court ruling; at least 57 died. Our 38-day reconstruction found the narrative peaked on the open-platform echo surface on 14 to 17 July, about two weeks before the surge, and decayed into the event. The same instrument reported what it did not find: no account-age clustering, shared content at 1.2 percent, and 8 manipulated media items among 131 scanned. One case does not make a typical figure. It does make the early-warning window the report asks for something that can be measured rather than asserted.
Reproducibility
A coordination scan re-run on a sealed corpus after 1,392 posts were removed returned all seven flagged pairs identical to four decimal places. The verdict depends on the behaviour of the accounts, not on how much was collected.
4. The horizons
2028. The report expects geoeconomic confrontation, misinformation and polarisation to be the three most severe risks, and half its respondents expect a turbulent or stormy two years. What we can do now, on the record: measure a narrative’s carriage on open platforms within a day of collection; give an early-warning window in days where one exists, and say so where it does not; publish nulls; and label a claim’s carriage without reading the claim, so that a finding travels as intelligence (STIX, DISARM) rather than as a moderation flag.
2031. Our planning assumption is that “is this real” becomes unanswerable from the artefact in the general case, and “who moved it, how, and with whom” is the only question with a deterministic answer. That requires scoring that is depth-stable and version-controlled (a verdict given in 2027 re-derivable in 2031 under the model of its day), verdicts that travel in the same channels as an indicator of compromise, and coverage that widens to closed platforms on licensed terms rather than scraped ones.
2036. The report asks governments to monitor “trust indicators and political volatility” the way they monitor financial markets for systemic exposure (p66). Today that means surveys. A narrative risk index derived from measured coordination, automation and velocity is a leading indicator rather than a lagging one, and it is falsifiable against what happened. If by 2036 there is a public record of measured cases, with the bounds stated, the report’s scenario in which “any scandal can be dismissed as a deepfake and any deepfake might be real” (p64) has at least one answer that is not an opinion.
5. What we do not do
We do not attribute operations to an actor. We do not read closed platforms without a licence. We do not judge truth; our labelling framework grades a claim’s standing under bounded retrieval and says so when it cannot. Our nulls are nulls for the methods and platforms used. Media authenticity is a signal from a third-party provider, not a verdict. One published investigation and one instrumented case carry most of the early-warning claim; the calibration set behind our coordination thresholds is small and we say so. The scale figures above were read on 15 October 2026 and cover every monitor on the platform, including our own research monitors.
6. What we will publish next
The lead-time series across every case, including zeros. Every coordination claim we test and fail to corroborate, with its controls. The labelling standard as a conformance specification other vendors can claim. And an annual re-mapping against each Global Risks Report, so that what we could measure in 2026 can be read against what we can measure in 2036.
By role: for CISOs, for chief risk officers and for government affairs.
