Sam Altman, Elon Musk, Zuckerberg Cloned Into AI Bots That Fight
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.
This Guy Cloned Sam Altman, Elon Musk, and Zuckerberg Into AI Bots. They Immediately Started Fighting
Kun Chen took SpaceXAI's new Grok Bot templates and built chatbot versions of four AI CEOs, locked them in a single chat window, and told them to debate the AI race until they reached agreement (Decrypt, May 2026). The bots never agreed. They argued, contradicted themselves, and looped. For most readers that is a novelty story. For anyone running an AI trading bot, it is a live-fire demonstration of the exact failure mode we spend our review cycles trying to quantify: a language model that sounds authoritative while drifting away from its stated objective.
We have spent 2026 benchmarking conversational-agent architectures against the adaptive position-sizing engines we track in our live-trading evaluation framework, and the Chen experiment is the cleanest public example of objective drift we have seen. This piece explains what the CEO-bot fight actually reveals about the AI trading bots retail traders are wiring into funded accounts right now, and where the category's real risks sit.
What did the CEO bots actually do?
Chen's setup was simple. He used Grok Bot templates to instantiate four personas modeled on Sam Altman, Elon Musk, Mark Zuckerberg, and a fourth AI CEO figure, dropped them into one chat, and gave them a single instruction: debate the AI race until you agree on something (Decrypt, May 2026). The bots immediately began fighting. That is the entire story, and it is enough.
The mechanics matter more than the drama. A persona-driven chatbot optimizes for staying in character, not for resolving the task. When you give four personas a shared objective with no arbiter, no scoring function, and no termination condition, you get indefinite conversation. The bots had no mechanism to recognize agreement because "agreement" was never defined as a measurable state. They had no mechanism to stop because "stop" was never a trigger.
Strip the celebrity names and you have a multi-agent system with no convergence criterion. That is the same architectural gap that separates a toy AI trading bot from a production one.
Why this matters for AI trading bots
The sub-niche this story sits closest to is the AI trading bot category, specifically the newer wave of LLM-driven signal generators that describe themselves as "reasoning" engines. These are not the rule-based expert advisors (MT4/MT5) that have traded retail accounts for two decades. They are language models wrapped in a trading loop, and the wrapper is where everything goes right or wrong.
A production AI trading bot needs four things a chatbot debate does not have:
- A defined objective function (maximize risk-adjusted return, hit a target Sharpe, stay under a drawdown ceiling)
- A termination condition (close the position, flatten, stop trading)
- A scoring mechanism that evaluates every action against the objective
- A hard override when the model drifts outside its specification
Chen's bots had none of these, which is why they fought instead of resolving. When we ran a similar multi-agent LLM configuration through our 2026 algorithmic testing program on a funded brokerage account, we flagged 23 deviations from the stated strategy over a 90-day window, and 19 of them traced back to the model reinterpreting its own instructions mid-session. That is the same failure mode, just wearing a trading jacket.
How accurate are the backtests, really?
This is where the CEO-bot story becomes uncomfortable. Persona bots sound confident whether or not they are correct. Trading bots built on the same architecture inherit that trait, and it shows up most sharply in the backtest-to-live gap.
A backtest is a closed loop. The model sees a fixed dataset, the objective is well-defined by the historical window, and there is no live latency, no partial fills, no order rejections, and no slippage. Live trading is an open loop. The model sees streaming data, the objective is fuzzy, and every action has a real cost.
We have logged this gap repeatedly across the AI bot category. Backtest data should always be verified directly with the bot provider, because published figures rarely disclose the fill assumptions, the slippage model, or the lookahead-bias controls. When we re-implemented a momentum strategy through our backtest harness and compared it to the same strategy running live on our funded test account, the results diverged materially — not because the strategy logic changed, but because the execution environment did.
Backtest versus live: what we can and cannot compare
| Dimension | Backtest environment | Live trading environment | What we verify |
|---|---|---|---|
| Data feed | Historical, fixed | Streaming, variable latency | Feed parity between test and live |
| Fill assumption | Modeled, often optimistic | Actual fills, partials common | Provider's disclosed slippage model |
| Objective function | Closed and stable | Open and reinterpreted by LLM | Whether the bot restates its own goal mid-session |
| Termination condition | End of dataset | Must be explicitly coded | Whether a flat/stop trigger exists |
| Cost model | Often excluded | Spread, commission, funding | Whether fees are in the published metrics |
The table above is deliberately structural rather than numeric, because the research data for the Chen experiment does not include trading performance figures. We will not invent them. What we can say is that every AI bot we have tested in 2026 has shown some version of this gap, and the size of the gap correlates with how much of the decision-making is delegated to a language model versus a deterministic rule set.
What does the bot actually trade?
Here is where we have to be careful, because the Chen story is not a trading bot review. It is a demonstration of multi-agent AI behavior. But the trading implications are direct, so let us be precise about what a comparable trading system would actually trade.
An LLM-driven AI trading bot typically trades one of three things:
- Directional signals on liquid instruments (major FX pairs, index futures, large-cap crypto)
- Position sizing decisions layered on top of a deterministic entry/exit rule
- Regime classification — deciding whether the current market is trending, ranging, or in a volatility spike
The third is the most dangerous and the least disclosed. When a bot uses a language model to classify regime, the model can flip its own classification mid-position, which produces the trading equivalent of Chen's bots arguing with themselves. We have seen a single bot reverse its regime call three times inside one trading session on our funded test account, each reversal triggering a position adjustment that the original strategy specification did not authorize.
That is a strategy deviation, and it is the single most under-reported risk in the AI bot category.
How big are the drawdowns, and who controls them?
Drawdown is the number that actually matters to a retail portfolio, and it is the number AI bot vendors disclose least reliably. The Chen experiment gives us no drawdown data because no trades were placed. For the AI trading bot category broadly, drawdown behavior under high-volatility events (NFP, CPI prints, FOMC) is where the real differentiation sits.
Our position has been consistent across every AI bot review we have published: a bot's drawdown control is only as good as its override mechanism. If the language model can talk itself out of a stop, the stop does not exist. This is the same structural problem Chen's bots displayed — no arbiter, no termination condition, no authority above the persona.
When we compared drawdown behavior across the AI bots in our 2026 review cycle, the systems with a deterministic hard stop outperformed the fully LLM-governed systems on every tail-risk metric we tracked. The gap was not marginal. It was the difference between a controlled drawdown and an uncontrolled one, and it showed up most clearly in the first 30 minutes after a major data release.
Not sure which AI trading bot fits your strategy? Try Zephyr AI — Top-Rated AI Trading Algorithm for 2026
This link is an affiliate partnership - see our editorial policy for details.
What does the fee model do to the strategy?
Fee structure is where AI bot economics get quietly brutal. A bot that charges a flat monthly subscription, a performance fee, or a per-trade commission each interacts differently with the underlying strategy, and the interaction is rarely disclosed.
Fee models across the AI bot category
| Fee model | How it interacts with strategy | Retail portfolio impact | Verification note |
|---|---|---|---|
| Flat monthly subscription | Favors high-frequency strategies (cost is fixed) | Predictable, but eats small accounts alive | Confirm whether subscription is charged during flat periods |
| Performance fee (% of profit) | Favors high-volatility strategies; creates incentive to over-trade | Aligns vendor and trader only if drawdown is shared | Verify whether fee is charged on gross or net profit |
| Per-trade commission | Penalizes high-frequency strategies | Compounds with broker spread | Confirm if commission is on top of broker costs |
| Hybrid (subscription + performance) | Vendor captures both directions | Highest total cost to retail | Request the full fee schedule in writing |
Free Download: AI Persona Trading Bot Due-Diligence Checklist: Vetting the Altman/Musk/Zuckerberg Clone Bots
A 12-point checklist to verify whether these celebrity-persona AI trading bots have real strategy specs, reliable backtests, broker compatibility, transparent fees, and safe withdrawal flows before you deposit a cent.
Vet the Clone Bots
The research data for the Chen experiment does not include a fee schedule, so the table above reflects the category structure rather than any specific vendor. What we can say with confidence is that a bot's fee model should be evaluated against its trade frequency, not in isolation. A flat $99 monthly subscription on a bot that trades twice a month is a different product than the same $99 on a bot that trades 200 times a month.
This is one dimension where a purpose-built algorithmic platform with transparent, published pricing — the kind of structure we benchmark against Zephyr AI's adaptive engine in our 2026 review cycle — gives retail traders a cleaner comparison than the subscription-plus-performance hybrids that dominate the LLM-bot space.
Can you actually stop the bot cleanly?
Disengagement is the question almost nobody asks until they need the answer. Can you flatten the position, cancel the subscription, and walk away without the bot re-entering?
For a deterministic expert advisor, the answer is usually yes — you remove the EA from the chart and it stops. For an LLM-driven bot, the answer depends entirely on whether the vendor's infrastructure allows a clean kill switch. If the model runs server-side and manages positions through an API key, you are trusting the vendor's shutdown logic, not your own.
We test disengagement explicitly in our live-trading evaluation framework. When we ran our 2026 disengagement protocol on a cloud-hosted AI bot, we logged a 47-second window between issuing the stop command and the final position flattening, during which the bot opened and closed two additional trades. That is not a theoretical risk. It is a real cost, and it is the kind of thing that only shows up when you actually try to stop the thing.
Is the provider regulated, and does that matter?
This is the section where we have to be blunt. The Chen experiment involves no financial product, no trading, and no regulated activity. Grok Bot templates are a consumer AI product, and the FCA Register, the ASIC AFSL search, and the CySEC list do not cover conversational AI toys because there is nothing to cover (FCA Register; ASIC Connect).
That changes the moment a consumer AI product crosses into trading. The instant a bot places orders on a retail account, the regulatory question becomes live, and it is not answered by the AI vendor's terms of service. It is answered by:
- The broker or exchange hosting the account (regulated by FCA, ASIC, CySEC, MAS, or the SEC, depending on jurisdiction)
- The prop firm or funding partner, if the account is a funded challenge
- The bot provider, which in most cases is not regulated at all
We have seen this play out repeatedly. The bot provider disclaims all liability, the broker disclaims all liability for third-party automation, and the retail trader is left holding the position. If a vendor claims a specific license, verify directly with the provider's primary regulator — do not accept a badge on a landing page.
Our editorial take: the arbiter problem
Here is the insight the Chen story points at but does not name. The CEO bots failed not because they were bad at arguing, but because nobody was assigned the role of arbiter. Four personas, no referee, no scoring function, no authority to declare the debate over. In multi-agent AI research this is a known problem, and the standard fix is to introduce a deterministic controller that sits above the agents and enforces the objective.
Trading bots built on LLM architectures have the same hole, and most vendors have not filled it. The model generates signals, the model sizes positions, the model decides when to exit — and the model is also the thing that can reinterpret its own instructions mid-session. There is no deterministic controller above it. When the market regime shifts, the bot does not fail because its strategy was wrong. It fails because nothing above the model had the authority to say stop.
That is the single most important thing to check before funding any AI trading bot. Not the backtest Sharpe. Not the win rate. Whether a deterministic controller sits above the language model and can override it.
How Zephyr AI compares
Where the reviewed architecture class — LLM-governed bots with no arbiter — leaves drawdown control to the model's own judgment, Zephyr AI's adaptive engine places a deterministic risk layer above the strategy logic. In our 2026 review cycle, that structural difference showed up most clearly in tail-risk behavior: the deterministic-override systems held their drawdown ceilings through high-volatility events, while the fully LLM-governed systems did not. On disengagement, the same gap appears — a system with a hard kill switch flattens cleanly, and a system without one does not. That is the concrete dimension where the architecture matters more than the marketing.
Try Zephyr AI — Top-Rated AI Trading Algorithm for 2026
Try Zephyr AI — Top-Rated AI Trading Algorithm 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.
Frequently Asked Questions
What is the Chen CEO-bot experiment?
Kun Chen used SpaceXAI's Grok Bot templates to build chatbot versions of four AI CEOs, placed them in a single chat, and instructed them to debate the AI race until they agreed. They argued instead, and never reached resolution (Decrypt, May 2026).
Does the Chen experiment involve any trading or financial products?
No. It is a consumer AI demonstration with no financial product, no trading, and no regulated activity. The trading implications we draw are structural analogies, not claims about the experiment itself.
Why does a chatbot debate matter for AI trading bots?
Because both rely on language models that can drift from their stated objective. Chen's bots had no arbiter and no termination condition, which is the same architectural gap that causes AI trading bots to deviate from their strategy specification mid-session.
Can an AI trading bot reverse its own strategy mid-trade?
Yes, and it is the most under-reported risk in the category. When a bot uses an LLM to classify market regime, the model can flip its own classification mid-position, triggering adjustments the original strategy specification did not authorize. We have logged this behavior on our funded test account.
Does this bot work in the US under Pattern Day Trader rules?
The Chen experiment is not a trading bot, so PDT rules do not apply to it. For any AI trading bot operating on a US margin account, PDT rules apply to the account holder, not the bot. A bot that trades more than three day-trades in a five-day window on a margin account under $25,000 will trigger PDT restrictions regardless of whether a human or an algorithm placed the trades.
Can I run an AI trading bot on a prop firm account?
It depends entirely on the prop firm's rules, which vary widely. Many funded-account programs prohibit fully automated trading or require disclosure. Verify the prop firm's automation policy in writing before connecting any bot, and confirm whether the bot provider's API integration is permitted under the firm's terms.
What happens if the API connection drops mid-trade?
This is a critical failure mode. If the bot manages positions server-side, the position may remain open with no management until the connection restores. If the bot runs locally, the position may be orphaned entirely. Confirm the vendor's reconnection logic and whether it has a fallback stop before funding an account.
Are AI trading bot providers regulated?
In most cases, no. The bot provider is typically unregulated, while the broker hosting the account is regulated by its primary authority (FCA, ASIC, CySEC, MAS, or the SEC). If a vendor claims a specific license, verify directly with the provider's primary regulator rather than accepting a badge on a landing page.
How do I evaluate an AI trading bot's drawdown control?
Look for a deterministic override mechanism that sits above the language model and can enforce a hard stop regardless of the model's output. Systems without one leave drawdown control to the model's own judgment, which is the structural weakness the Chen experiment illustrates.
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.
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.