Artificial intelligence is often presented as either the solution to misinformation or its main cause. The honest answer is that it is a powerful tool with specific strengths and specific weaknesses, and a responsible system has to be designed around both.

What AI does well

  • Breaking text into claims. A long article can contain a dozen separate factual assertions mixed with opinion. Models are good at isolating the parts that can be checked.
  • Working across languages. A claim in Persian, a source in Russian and a study in English can be compared in one pass.
  • Spotting framing. Loaded verbs, missing qualifiers and inflated numbers follow recognisable patterns.
  • Speed. A first analysis in under a minute, instead of hours.

What AI does badly

  • Knowing what happened recently. Models are trained on data up to a certain date. Events after that are unknown to them unless the information is supplied.
  • Admitting ignorance. Models are trained to produce fluent answers. Without careful design, they fill gaps with plausible inventions.
  • Citing sources reliably. A model can produce a realistic-looking reference or web address that does not exist.
  • Taking responsibility. A model cannot be held accountable. People and institutions can.

Designing around the weaknesses

Honest Lens does not trust the model's output blindly. Several safeguards are built into the Claim to Evidence engine:

  • Every cited link is checked by the server. Links that do not resolve are removed and marked, not shown as evidence.
  • "Insufficient Evidence" is a correct answer. The model is told explicitly that guessing is a failure.
  • No invented addresses. If the exact address of a document is not known, the source is named without a link.
  • The model is named. Every report says which AI model produced it, so readers know what they are relying on.
  • No trust scores. The output is evidence and reasoning a reader can inspect, not a number to accept.

The right role for AI

The most useful role for AI in fact-checking is not judge but assistant: a fast first pass that organises claims, surfaces evidence and flags what needs a human look. The final judgement, especially on sensitive or high-stakes claims, belongs to people who can be held responsible for it.