Google has introduced a ‘Verified Sources’ mode for its Gemini assistant that declines to answer a query unless it can point to a real, retrievable document — a notable break from the fluency-first design that has defined consumer chatbots since 2023. The change follows a bruising run of incidents in which UK solicitors were sanctioned for filing court submissions citing case law that never existed, much of it traced back to generative AI tools. For a technology sold partly on its confident command of language, the move amounts to an admission: sounding right is no longer good enough.
The feature, rolling out first to Gemini’s paid and enterprise tiers before a wider release, constrains the model to claims it can tie to a source in an indexed corpus. Where it cannot, it says so. Google frames this as a deliberate shift toward ‘auditable’ answers — a phrase that, until recently, sat awkwardly alongside the marketing of large language models as effortless, all-knowing oracles.
The scandal that forced the issue
British courts have spent the past 18 months grappling with a peculiarly modern problem: lawyers submitting authorities that simply do not exist. In several cases before the High Court and tribunals, judges found that citations — complete with plausible-looking party names, neutral citation numbers and even fabricated quotations — had been generated wholesale by AI tools and never checked against a real reporter.
The professional consequences have been severe. Solicitors have faced wasted-costs orders, referrals to the Solicitors Regulation Authority and public reprimands from the bench. The Bar Council and Law Society both issued guidance warning members that the duty to verify rests with the practitioner, not the machine — but the reputational damage to AI vendors courting the legal market was already done.
“The legal sector was supposed to be a flagship enterprise market for these tools, and instead it became the cautionary tale,” said Dr Priya Nandakumar, a researcher in computational law at the fictional Institute for Digital Governance. “You cannot sell a research assistant to a profession built on precedent if it invents the precedent.”
How ‘Verified Sources’ actually works
Rather than generating an answer and then attempting to find citations after the fact — the approach behind many bolt-on ‘grounding’ features — Verified Sources reverses the order. The system first retrieves candidate documents from an indexed set, then constrains the model’s output to assertions it can attribute to those documents. If retrieval returns nothing relevant, the assistant returns an explicit ‘no verifiable source found’ rather than improvising.
Google says the mode supports custom corpora, allowing law firms, financial institutions and public bodies to point Gemini at their own document stores — case management systems, internal knowledge bases or licensed legal databases. Each factual claim is rendered with an inline link to the underlying passage.
- Refusal over invention: the model abstains when it cannot ground a claim.
- Inline attribution: every assertion links to a retrievable source passage.
- Custom corpora: organisations can restrict answers to vetted internal documents.
The trade-off is candour about limits. Early testers report that Gemini in this mode answers fewer questions and hedges more often — a deliberate cost that Google is, for once, willing to advertise.
Raising the bar for rivals
The competitive signal is hard to miss. OpenAI, Anthropic and a clutch of specialist legal-AI startups such as Harvey and Lexis+ AI are all chasing professional users who need answers they can defend in front of a regulator or a judge. By making verifiability a headline feature rather than a footnote, Google is attempting to reset the criteria on which these products are judged.
“Fluency was the differentiator in 2023. Auditability is the differentiator now,” said Marcus Feld, principal analyst at the fictional firm Northgate Research. “Whoever convinces compliance officers that their model knows what it doesn’t know wins the enterprise contract. That’s a very different race.”
Sceptics caution that the feature is only as trustworthy as the corpus behind it. A model that faithfully cites a poorly maintained or biased database will produce confident, well-linked errors. There is also the risk of ‘verification theatre’ — users assuming a citation guarantees accuracy without clicking through to check that the source actually supports the claim.
The pivot away from confident guessing
For all its limits, Verified Sources reflects a maturing of the industry’s self-understanding. The hallucination problem was long treated as a bug to be engineered away; increasingly it is recognised as an intrinsic property of probabilistic text generation, to be contained rather than cured. Constraining a model to retrievable evidence is less glamorous than promising omniscience, but it is honest about what these systems are.
What this means: Google’s Verified Sources mode marks a quiet but consequential shift in how AI assistants are sold to people who cannot afford to be wrong. For UK solicitors and other regulated professionals, it offers a tool whose answers can, in principle, be checked rather than merely trusted — though the burden of verification still rests squarely with the human. For Google’s rivals, it sets a new benchmark: in the professional market, the most valuable thing an AI can do may be to admit when it does not know.
Photo by Markus Winkler on Pexels
