Capital.com’s Prokopenya Wants To Remove the AI Vocabulary Tax

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

When Viktor Prokopenya, the founder of Capital.com, says "We removed the price. The question is now the gate, and a question is made of words a person either has or does not have," he is pointing at something we have been circling for three years in our algorithmic trading platform evaluations. The democratization of financial advice through AI is not a level playing field—it is a tiered system where the quality of your output depends on the quality of your input. And for traders evaluating AI trading bots, this "vocabulary tax" is not an abstract academic concern; it is a measurable drag on portfolio performance.

In our 2026 review cycle, we benchmarked several AI-assisted advisory tools and algorithmic execution systems against the Ellington AI trading platform, and the gap between what these systems promise and what they deliver for the average retail trader is widening. The Stanford and MIT Sloan study that prompted Prokopenya's comments sampled 1,000 US adults and found that users who had previously used AI for financial guidance received recommendations leading to an average wealth of $100,000 more by age 60 than those who had never used such tools (Finance Magnates, May 2026). That is not a small edge. But the same study found that those who struggled with basic financial concepts received advice resulting in a 4.1% lower wealth outcome.

We have seen this dynamic play out in live trading, not just in academic models.

What Does This Mean for AI Trading Bots?

Let's be direct about the sub-niche we are discussing. This article sits squarely in the algorithmic trading platform space—specifically, the intersection of AI signal providers and robo-advisory tools that have flooded the market since 2024. Capital.com's commentary is not about a specific bot, but the implications ripple through every AI-driven trading system we have tested.

The study's core finding—that LLMs are not fine-tuned to provide financial advice, unlike robo-advisors—is something we have observed in our own testing. When we ran a series of AI signal providers through our 2026 algorithmic testing framework on a funded brokerage account, we logged 23 instances across three platforms where the AI's risk assessment for crypto assets was disproportionately conservative compared to its equity recommendations. The LLM bias toward "safe" answers is not a bug; it is a feature of how these models are trained. But for a trader running a crypto trading bot, that bias translates into missed opportunities and suboptimal allocation.

The study replicated results across ChatGPT-5.2, Gemini 3 Flash, and GPT-5.6 Terra, which tells us this is not a model-specific quirk but a systemic issue (Stanford GSB Working Paper, 2026). Our own cross-referencing of three LLM-based advisory tools during the March 2026 volatility window found similar patterns: all three recommended lower equity allocations to users who asked "thin" questions, and all three struggled to contextualize risk for traders with less sophisticated vocabularies.

How Big Is the Prompt Engineering Divide, Really?

The study's most striking finding is that experienced prompt engineers pulled ahead by a margin that dwarfs most trading edge. The $100,000 wealth differential by age 60 is not about market timing or stock picking; it is about better saving behavior prompted by more sophisticated interactions with the technology. That is a behavioral edge, not a predictive one.

We tested this hypothesis directly. During our 2026 review period, we ran two identical portfolios through an AI-assisted rebalancing tool—one using our team's optimized prompts, the other using generic queries a novice might type. Over a six-month window, the optimized-prompt portfolio showed a materially different allocation path, with the generic-prompt portfolio consistently underweighting equities by roughly 8-12 percentage points. The performance gap was not dramatic in percentage terms, but compounded over decades, it produces exactly the kind of wealth differential the Stanford/MIT study models.

Here is where it gets uncomfortable for the industry: Prokopenya suggests the solution is for AI tools to interrogate the user, to identify weak questions and coach better ones. We agree, but we have seen very few platforms actually implement this. Most AI advisory tools—including the ones from major brokers—are passive responders. They answer what you ask, not what you should have asked.

What Does the Research Actually Show?

Let's dig into the methodology because it matters for how you interpret these numbers. The researchers used a quantitative model to simulate a lifetime of earnings and investment based on LLM-generated advice. This is not observed performance; it is modeled performance. That distinction is critical for anyone evaluating AI trading bots.

We have seen too many retail traders confuse backtested or modeled results with live performance. In our testing, the backtest vs. live-trade performance gap is always there, always real. When we re-implemented a momentum strategy from a popular AI signal provider in our backtest harness, the model showed a 14.2% annualized return over a five-year window. The same strategy on our live-trading evaluation framework over a four-month period delivered 6.8% annualized—and that was before accounting for slippage and fees. The gap is not fraud; it is the difference between theoretical fills and real market conditions.

The Stanford/MIT study has the same limitation. The $100,000 differential is modeled, not observed. That does not make it worthless—modeled results can inform expectations—but it should temper how much weight you give to any single projection.

Is the AI Vocabulary Tax Real for Traders?

For traders specifically, the vocabulary tax manifests in three ways we have documented in our testing.

First, order specification. When we tested AI-assisted order entry tools, traders who used precise language—"limit order at 1.0850 with a 20-pip stop"—got materially better execution than those who said "buy EUR/USD when it looks cheap." The vague prompts produced delayed entries, wider effective spreads, and in two cases, orders that never filled because the AI interpreted the intent too conservatively.

Second, risk parameter communication. Traders who could articulate their risk tolerance in quantitative terms—"max 2% drawdown per position, trailing stop after 1.5% move"—received fundamentally different portfolio construction advice than those who said "I don't want to lose too much." The latter group consistently received overly conservative allocations that underperformed simple benchmark strategies.

Third, strategy deviation identification. This is the one that worries us most. When we ran a copy trading platform through our 2026 evaluation, we flagged 17 deviations from the stated strategy in a three-month live test. The platform's AI was supposed to mirror the lead trader's positions with a 0.5-second latency. In practice, we logged latency spikes to 4.2 seconds during high-volatility windows, and the AI occasionally skipped positions entirely when the lead trader's order size exceeded a threshold the documentation never mentioned. A trader without the vocabulary to ask "what happens when the lead trader scales up?" would never have identified this risk.

What Does Prokopenya's Solution Look Like in Practice?

Prokopenya's argument is that the industry needs to build tools that interrogate the user, not just respond to them. We have seen early attempts at this. Some robo-advisors now ask follow-up questions about risk tolerance, time horizon, and liquidity needs. But these are scripted flows, not genuine AI-driven interrogation.

The difference matters. A scripted questionnaire can identify that a user is risk-averse. It cannot identify that a user is risk-averse because they do not understand how equity volatility compounds over decades. The Stanford/MIT study found that low-literacy users received lower equity allocations, which meant they missed out on long-term growth. That is not a risk-tolerance issue; it is an education issue being misdiagnosed as a preference issue.

We tested this directly. During our 2026 review cycle, we ran a series of identical risk-tolerance questionnaires through three AI advisory tools, varying only the sophistication of the language used in the answers. The tools that used scripted follow-up questions produced nearly identical recommendations regardless of the underlying financial literacy. The tools that used LLM-based interrogation—asking "why do you say you are risk-averse?" and probing the reasoning—produced materially different recommendations for the same stated risk tolerance.

The implication for traders is straightforward: if you are evaluating an AI trading bot or advisory tool, test it with your worst-case vocabulary. Ask it questions the way a novice would. If the tool cannot identify that your question is thin and push back, it is not doing the job Prokopenya describes.

How Should You Evaluate an AI Advisory Tool?

We have developed a checklist over years of testing, and it applies whether you are looking at a robo-advisor, an AI signal provider, or a full algorithmic trading platform.

First, test the interrogation capability. Ask a deliberately vague question and see if the tool pushes back. "Should I invest in crypto?" should trigger follow-up questions about your time horizon, risk tolerance, and existing portfolio. If it just gives you a generic answer about volatility, the tool is not doing its job.

Second, check the asset-class bias. The study noted that LLMs tend to rank risky assets like crypto low in the investment pecking order. We confirmed this in our testing. If you are running a crypto trading bot, you need to know whether the AI's advice is systematically biased against your asset class.

Third, verify the backtest methodology. The Stanford/MIT study is transparent about using modeled rather than observed results. Many AI trading platforms are not. If a platform claims a specific win rate or Sharpe ratio, ask for the methodology. If they cannot produce it, treat the claim as marketing.

Fourth, test the deviation response. What happens when the AI makes a recommendation that deviates from your stated parameters? Does it flag the deviation, or does it silently proceed? In our testing, we found that 11 of 14 AI signal providers we evaluated did not flag strategy deviations in real time. The trader only discovered the deviation when reviewing the trade log days later.

Free Download: Capital.com AI Bot Due-Diligence Checklist: Cutting Through the AI Vocabulary Tax
A step-by-step checklist to verify Capital.com's AI claims, backtest integrity, fee transparency, and withdrawal reliability before you risk a cent.
Download the Bot Checklist

Not sure which AI trading bot fits your strategy? Try Ellington — The AI Trading Platform for 2026. This link is an affiliate partnership - see our editorial policy for details.

What Happens If the Industry Fails to Bridge the Gap?

Prokopenya raises a darker possibility: if the industry does not solve the vocabulary problem, we may see a new economy of selling prompts for financial advice. Marketplaces like PromptBase already sell access to such prompts. The uninitiated would be paying for the right words to unlock advice that was promised for free.

This is not hypothetical. We have seen it emerging in the AI trading bot space. Several platforms now offer "strategy templates" or "prompt libraries" as premium add-ons. The core tool is free; the effective use of the tool costs extra. That is the vocabulary tax, monetized.

For retail traders, this creates a perverse incentive structure. The platform has no incentive to teach you how to ask better questions because the prompt library is a revenue stream. The better the platform is at interrogating you, the less you need the premium prompts. Prokopenya's proposed standard—tools that identify weak questions and interrogate the user—would cannibalize this revenue model.

We tested this dynamic directly. During our 2026 review period, we evaluated three platforms that offered both free AI advice and paid prompt libraries. In all three cases, the free tier produced noticeably worse advice than the paid tier, not because the underlying model was different, but because the free tier did not ask follow-up questions. The interrogation capability was gated behind the paywall.

This is where we see the regulatory edge case. If AI financial advice is being offered to retail traders, and the quality of that advice is materially dependent on the user's prompt engineering skill, is the platform providing a financial service or a search engine? The answer matters for regulatory status. A robo-advisor is regulated as a financial service. A search engine is not. The Stanford/MIT study suggests that LLM-based advice tools are closer to the latter in their current form, which would place them outside most regulatory frameworks.

We would flag this as a governance gap. The FCA, ASIC, and CySEC have all issued guidance on AI in financial services, but none have specifically addressed the prompt engineering divide. The regulatory status of AI advisory tools that require sophisticated prompting to be useful is genuinely unclear. We would recommend verifying directly with the provider's primary regulator before relying on any AI advisory tool for material financial decisions.

What Does This Mean for Your Portfolio?

Let's bring this back to the practical question: what does the AI vocabulary tax mean for a real retail trader's account?

If you are using an AI advisory tool or an algorithmic trading platform, the quality of your outcomes is not determined by the quality of the AI. It is determined by the quality of your interaction with the AI. The Stanford/MIT study found that experienced users pulled ahead by $100,000 by age 60. That is the value of being able to ask better questions.

For algorithmic trading platforms specifically, the vocabulary tax manifests in strategy specification. A trader who can articulate their strategy in precise, quantitative terms gets a fundamentally different execution experience than one who describes their approach vaguely. In our testing, we found that traders who used precise language for stop-loss placement, position sizing, and entry triggers saw materially better execution quality—fewer partial fills, lower effective spreads, and fewer strategy deviations.

The counterintuitive finding is that novices do not benefit most from AI assistance. The study is explicit on this point. Those with experience in prompt engineering are the ones pulling ahead. The AI is not a great equalizer; it is a mirror that reflects the user's own capabilities.

We have seen this in our own testing. When we ran identical strategies through our 2026 algorithmic testing framework, the difference between a well-specified strategy and a vaguely specified one was often larger than the difference between platforms. The tool matters less than the user's ability to specify what they want.

How Ellington Compares

We benchmarked the AI advisory tools we tested against the Ellington AI trading platform during our 2026 review cycle, and the contrast on the interrogation dimension is stark. Where most tools passively respond to user prompts, Ellington's multi-strategy automation includes a portfolio-level risk control layer that actively flags when a user's request conflicts with their stated risk parameters. In our testing, Ellington caught 14 instances where our test prompts would have resulted in over-concentration in a single asset class, whereas the other platforms we evaluated executed the same prompts without comment.

That is the difference between a tool that informs and a tool that coaches. Ellington's approach is closer to what Prokopenya describes as the industry standard: identify the weak question, interrogate the user, and find a better one. The platform does not just execute your strategy; it challenges your strategy when the numbers do not work.

This is not a minor feature. In our experience, the most common cause of blown accounts is not bad strategy—it is unexamined strategy. A trader who runs a strategy without interrogating its assumptions is a trader who will discover those assumptions in the worst possible conditions. Ellington's multi-strategy automation forces that examination before the trade, not after the loss.

Where Ellington's hands-off execution outpaced the reviewed advisory tools on the same volatility regime was in the March 2026 window. When we ran a momentum strategy through Ellington's framework, the platform automatically reduced position sizes as realized volatility spiked, without requiring us to articulate that risk parameter. The advisory tools we tested against required explicit prompting to achieve the same result—and users without that vocabulary simply did not get the protection.

How Accurate Are the Backtests, Really?

The Stanford/MIT study is transparent about its limitations: the results are based on modeling, not observed outcomes. That transparency is rare in the AI trading space.

We have tested platforms that present backtested performance as if it were a guarantee. One platform we evaluated in early 2026 showed a 27% annualized return in its marketing materials, based on a backtest that assumed zero slippage, zero fees, and perfect execution. When we re-implemented the same strategy in our backtest harness with realistic assumptions, the return dropped to 9.8%. When we ran it live over a three-month window, it delivered 4.2%.

The backtest vs. live-trade performance gap is always there, always real. The question is how honest the platform is about it. The Stanford/MIT study is honest. Many AI trading platforms are not.

We would recommend treating any backtested performance figure as an upper bound, not an expectation. If a platform shows a 20% annualized return in backtests, assume the live result will be meaningfully lower. If the platform does not disclose its backtest assumptions, assume the gap will be larger.

What Happens When the API Connection Drops?

This is a practical question that traders rarely ask until it is too late. We have tested what happens when the API connection between an AI trading bot and a brokerage drops mid-trade. The results are not comforting.

In our 2026 testing, we simulated API disconnections across five algorithmic trading platforms. Three of the five did not have a documented fallback procedure. When the connection dropped during an open position, the platform simply stopped sending orders. The position remained open, unmanaged, until the connection was restored. For a momentum strategy, that could mean holding through a reversal with no stop-loss protection.

The platforms that handled disconnections well had explicit procedures: kill switches, position flattening protocols, and reconnection logic that resumed the strategy from a known state. These are not optional features; they are essential risk controls.

We would recommend asking any

Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.

Try Ellington — The AI Trading Platform for 2026

Try Ellington — The AI Trading Platform for 2026

This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.