In October 2025, Deloitte agreed to partially refund the Australian government for a report riddled with fabricated citations and a fake quote attributed to a federal court judgment. The document had been produced with AI assistance, and it read exactly like every other confident, professionally worded consulting report. Nobody caught the errors by tone. Someone had to catch them by checking, as Fortune reported at the time.
That is the uncomfortable truth sitting underneath most AI adoption stories right now, and it is worth exploring alongside British Wire’s technology coverage of where AI is genuinely earning its place in business. AI systems do not sound less certain when they are wrong. If anything, they sound more certain, because they are built to produce fluent, plausible text rather than verified fact. A confident sentence and a correct sentence look identical until someone checks.
Why Confidence Isn’t the Same as Accuracy
The instinct most businesses have is to treat a well-written AI answer as a trustworthy one. That instinct is exactly backwards. A model that hallucinates a statistic, a legal citation, or a translated clause does not flag its own uncertainty. It writes the wrong answer with the same tone as the right one. Translation is a good illustration of how this problem is evolving: surface-level errors have largely disappeared, and what’s replaced them is subtler and harder to catch, namely errors that read fluently but say something subtly different from what was intended. That is what makes single-model AI output risky in high-stakes settings: finance, legal research, compliance documents, and any customer-facing content where a mistake becomes public.
The fix is not asking the AI to double-check itself, since a model rechecking its own reasoning tends to agree with its own mistake. The fix is a structural one: build in a way to catch disagreement before it reaches a decision.
The Three-Question Test
Before trusting any AI output that matters, run it through three checks.
- Would this answer survive being asked twice, differently?
Rephrase the same question and see whether the facts, figures, or recommendation hold steady. If the answer shifts depending on how it was asked, that instability is a signal, not a coincidence.
- Can it show its work, or just its answer?
Ask the model to point to where a specific claim came from. If it cannot produce a real, checkable source, or the source doesn’t actually say what’s being claimed, the output is manufactured plausibility dressed up as fact.
- What happens when you check it against an independent second opinion?
Run the same query through a different system entirely. Agreement is a reasonable, though imperfect, signal of reliability. Disagreement is the most useful data point in the whole exercise, because it tells you precisely where a human needs to step in before anything ships.
That third question is really the whole point. Recruiters already understand this instinct in a different context: AI can screen and rank candidates faster than any human, but the strongest hiring processes keep a human in the loop rather than letting the model make the final call alone. The same principle applies anywhere AI output feeds into a business decision. AI should narrow the work, not replace the judgment.
The Root Cause: One Model, One Opinion
Here is the part most companies miss. The verification burden that comes with AI adoption isn’t really a training problem or a prompting problem. It’s an architecture problem. A single AI model is a single point of failure, and businesses have already learned this lesson in a completely different domain. When supply chains leaned on one supplier or one region for critical materials, relying on a single source turned out to be a risk businesses could no longer afford once conditions shifted, and diversification became the obvious fix. The same logic applies to AI. As Deloitte’s own research into enterprise AI governance points out, organisations that scale AI responsibly build in independent validation rather than trusting a single system’s output outright. One model, however capable, is still one opinion.
What Consensus Checking Looks Like in Practice
This is where cross-model consensus checking earns its place, and it doesn’t require a research team to implement. The principle is simple: instead of taking one model’s answer at face value, compare outputs from multiple models on the same task and treat disagreement as the signal to slow down.
Translation is a useful, low-drama example of exactly this problem, because a wrong answer there is easy to spot in hindsight and expensive to miss in the moment, and it rarely announces itself. A mistranslated clause, a shifted number, or a misjudged tone can cause real damage while still reading perfectly fluently. MachineTranslation.com’s SMART consensus system takes this approach a step further, running every query through 22 different AI models simultaneously and selecting the output where those models actually agree, rather than trusting whichever single model happens to answer first. It’s the same three-question test applied automatically at the point of output, not bolted on afterward as a manual review step.
The Takeaway
The businesses avoiding costly AI mistakes in 2026 are not necessarily using better AI models. They are using more than one, and they’ve built a habit of checking before they act on an answer that sounds too smooth to question. The three-question test won’t catch everything, but it catches the failure mode that matters most: the wrong answer that never had to sound wrong to do damage.
