Espen Skogen has spent two decades inside the delivery machinery of global finance, and a pattern has emerged across his observations that few executives want to hear. AI handles implementation, testing, first drafts, and research synthesis with remarkable fluency. It does all of this while remaining completely indifferent to whether the output is correct. The combination of fluency and indifference produces something specific: confident output that looks finished and reads as authoritative, whether or not it is true.
Without insight, that combination is downright dangerous.
The danger is measurable. As of April 2026, researchers have documented 1,313 court proceedings worldwide containing AI-fabricated content, with 496 of them involving licensed attorneys. Individual sanctions have climbed from $5,000 in 2023 to $55,597 in 2025, an elevenfold escalation in two years. These are professionals, trained to verify sources, submitting citations that never existed.
Skogen's reading of this data is characteristically structural. The failure sits upstream of the tool.
The legal system offers the cleanest demonstration of the mechanism, which is why Skogen keeps returning to it. As a qualified solicitor who also architects trading systems, he watches both domains absorb the same technology with the same blind spot.
Tools like ChatGPT have given litigants in person access to court proceedings at a level unheard of five years ago. They can draft submissions, cite precedent, and structure arguments in the language of the court. On the surface, this looks like democratization.
The problem lives underneath the surface.
Stanford research found error rates of 69 to 88 percent for general-purpose language models on legal queries. Even specialized legal AI tools showed error rates between 17 and 34 percent. When the model fabricates a case, it does so in perfect legal register, with plausible citation formatting and a confident tone. A trained lawyer catches the fabrication because the fabrication conflicts with years of accumulated pattern recognition. The litigant in person has no such pattern library.
They don't know what they don't know.
💡 The value of expertise was never the ability to produce a document. It was the ability to recognize when a document is wrong.
Here is where Skogen's analysis diverges from the standard AI commentary, and where it becomes uncomfortable for the institutions he works with.
The specific tasks AI now absorbs, the first-draft generation, the research synthesis, the document review, are precisely the tasks that historically taught junior professionals what good looks like. In consulting, an estimated 30 percent of tasks face automation, concentrated heavily in the junior work of data analysis and report preparation. That work was never valuable primarily for its output. It was valuable as a training ground for judgment.
You learned what a good risk report looked like by writing forty mediocre ones and having each one corrected. You learned what clean architecture felt like by building something ugly and living with the consequences. The repetition built the pattern library that later allowed you to glance at an output and sense that something is off before you can articulate why.
Remove the repetition and the pattern library never forms.
The hiring data already reflects this. Analysis suggests AI could eliminate 56 percent of entry-level jobs within five years, with junior technical support hiring down 15 percent while senior professionals in the same fields maintain stable employment. Firms are, in effect, spending down an inheritance. The senior people who can validate AI output today built their judgment on the very tasks the AI now performs.
With fewer people gaining basic experience, firms slim down their future talent pool and compress the base layer of their talent pyramid, which over time impacts upper tiers and narrows the available pool for mid-level and leadership roles.
In finance terms, this is the one analogy Skogen permits himself: institutions are liquidating an asset that took decades to appreciate, and booking the sale as a productivity gain.
Skogen operates at the intersection of law, technology, and financial services, and that vantage point makes the sector-specific risk hard to ignore. The Bank for International Settlements has warned that wider AI adoption without corresponding expertise could result in insufficient understanding and ineffective management of risks to financial institutions and the financial system itself.
Consider what that means in practice on a trading floor or in a risk function. A model confidently explains a margin shift. The explanation is fluent, internally consistent, and wrong, because the model filled the gap with probability rather than truth. If the person reviewing it has never manually decomposed a margin calculation, the error passes. It compounds. It informs a decision.
A 2026 expert study identified finance as one of the sectors most vulnerable to AI-driven risks, with 18 of 24 identified risks carrying more than a 10 percent probability of catastrophic outcomes exceeding $100 billion in losses within five years. Those numbers reflect a system where verification capacity is thinning at exactly the moment output volume is exploding.
⚠️ The hallucination problem worsens in more advanced models despite mitigation efforts.The better the model sounds, the more expertise is required to catch it when it lies.
Skogen built RocketFin on a conviction he has held long before language models entered the conversation: if you think you have a technology problem, you've misdiagnosed it. It's a people problem wearing a technical mask.
AI adoption follows the same law. The organizations getting hurt treated the tool as a substitute for judgment. The organizations extracting genuine value treated it as an amplifier for judgment that already existed.
His delivery model at RocketFin makes the logic explicit. Small teams of senior specialists, people who have walked the walk, who have built the trading platform and read the contract and shipped the risk system, use AI aggressively for implementation and testing. It works because every output lands in front of someone who knows what good looks like. The AI accelerates the work. The senior practitioner validates it. The sequence matters, and it only functions because the human in the loop earned their pattern library the slow way.
Competence compounds faster than coordination overhead. It also compounds faster than compute.
The practical implications follow directly from the evidence chain.
Skogen's position resists both camps in the current debate. The tools are genuinely useful, and the productivity gains in implementation and testing are real. The danger enters when access to output gets mistaken for possession of insight, because the litigant in person and the under-trained analyst share the same vulnerability: when the machine lies to them confidently, they lack the reference points to see it.
Insight remains the scarce resource. Everything AI produces is a claim awaiting validation, and validation requires someone who has done the work the AI now does.
So the question worth sitting with is this: when your AI gives you a confident answer next quarter, who in your organization still knows enough to tell you it's wrong?
https://www.haqq.ai/blog/when-ai-lies-to-the-court