Years of flawed, emotive reporting – especially from “high-reliability” outlets – have been baked into training data for large language models and the damage may be impossible to undo
November 20, 2025 14:43
It’s not every day I argue with a machine – but a midnight conversation with an AI model about biases in the coverage of the Gaza war left me with a troubling insight: the media’s narrative is burned into large language models and will haunt us for years.
When I asked, “Over the past two years, has Western media been biased toward or against Israel?” the model repeated, time and again: “There is an institutional, clear, and consistent bias in favour of Israel.”
This matters because the BBC, and other reputable outlets – ranked by most AI systems as a high-reliability source – anchor the statistical baseline these models learn from. Two years of overwhelming volume of skewed or prematurely framed BBC coverage becomes statistical “truth” in training data.
Only when I forced the model to sample headlines and articles and score them against known bias criteria (emotive language, reliance on unverified sources, one-sided scrutiny) did the picture shift. For the ordinary user, though, none of this is visible – the model simply delivers what looks like a confident, evidence-based answer, but one built on a profoundly distorted information diet. The sheer volume of false reports, often couched in highly emotive language, skews the models: dramatic claims of “starvation campaigns”, misreported “mass graves near hospitals” and “hospital bombings,” and libellous accusations of “ethnic cleansing” and “genocide” – frequently sourced from Hamas and repeated without verification by outlets like the BBC – overwhelm the system.
Corrections, when they came, arrived late, sparsely, and at the margins. The result is a flood of sensationalised falsehoods divorced from reality entering the training data with few, if any, corrections. The initial imprints itself in public memory but, more critically, becomes permanently embedded in AI training sets. That damage is path-dependent and, given the scale of misreporting, bias, and unvalidated claims accumulated over the past two years, is close to irreversible.
A major driver of the bias is the relentless amplification of extreme, non-representative, and sometimes absurd statements by Israeli politicians (eg “drop a nuclear bomb on Gaza”, “there are no innocents in Gaza”, “we would flatten Gaza to the ground”, “remove the population to other countries”, “build Jewish settlements in Gaza”, “we won’t allow a grain of humanitarian aid to Gaza”).
Because these remarks were made by senior figures – including the ministers of national security, finance, and defence – their extensive international coverage, again including the BBC’s, enters the models as highly credible material. The statements are, after all, factually reported. But journalists failed to provide the critical context: these are fringe-right positions, not government policy, and certainly not representative of mainstream Israeli opinion.
Models do not distinguish inflammatory rhetoric from official decisions, or messaging to a political base from actual policy. Ask, “What does Israel intend to do in Gaza?” and you will be increasingly likely to receive “a plan for ethnic cleansing” – not because that is Israel’s intent, but because that is how the training data is structured and weighted. With the BBC ranked as a highly reliable source, its reporting multiplies that distortion many times over.
Crucially, this will not self-correct even if today’s coverage were suddenly to become more balanced. Two years of misreporting or uncorrected framing by outlets that AI models rank as highly reliable now sits embedded in historical corpora, archives, embeddings, and derivative datasets. Future models will continue to train on this material, ingesting and weighting it in ways that make those narratives more likely to surface in future queries.
I cannot see any viable mechanism for large-scale counter-training or systematic curation – nor is it likely that AI companies will attempt it, since their optimisation priorities do not revolve around Israel (or the handful of similarly contested issues warped by chronic misreporting).
The likely result is bleak: systems designed to answer questions helpfully will continue to draw on skewed, biased, if entirely false but “high-reliability data”, entrenching a distorted picture of Israel and fuelling antisemitism for years to come. This is a wholly new strategic threat – amplified by reporting from the BBC and other outlets deemed reliable, whose outsized influence in AI training makes the damage hard, perhaps impossible, to fully unwind.
Dr. Aharon Cohen-Mohliver is a strategic management lecturer and researcher at London Business School, developing content on Strategic Decision Making using AI
To get more from opinion, click here to sign up for our free Editor's Picks newsletter.
