My AI Post About Your AI PostBitcoin / U.S. dollarBITSTAMP:BTCUSDholeyprofitThis isn't a post to hate on AI or even those who are taking advantage of it to do the automations. If there can be rewards for low effort people are going to do that. Instead its an appeal to people out there who may want good content. If it becomes the ingrained norm that low effort content is ranked equal to, or higher than, high effort content -no one is going to make an effort. And this site is going to become an endless stream of narrative based posts, thin TA and a comment section of bland agreement. A high percentage of the posts on the top ideas pages are no/low effort LLM content and what is more worrying about this is this is driven there by engagement metrics and over 70% of the comments are clearly botted. End to end botted. No one read the post. It was scraped, sent to a LLM with a prompt like "Engage positively with this content". And ITS BAD. My broker is running a misconfigured end to end LLM bot to comment here. They have not even got it good enough to get the context of the post its commenting on right. IF BOT POSTS DOMINATE THE SITE'S RANKING WITH BOTTED REPLIES ITS NOT WORTH HUMANS COMPETING. So, support human writers if you want them. With all due respect to LLMs - we still write considerably better trading analysis. Well, you know - case to case basis. ================================================================================== With all that said, here is my fully disclosed AI dump. I posted top ideas posts and comments sections to GPT with neutral prompts to assess the odds these were end to end botted systems: ================================================================================== Looking at the actual posts rather than just the general page, there is a much clearer pattern. I would not treat this as a binary “LLM / human” classification. The useful question is: How much actual cognitive/analytical work does the post appear to require from an LLM? The really interesting finding There are two different kinds of LLM posts here. Group A — “write me a trading post” These are the ones I would classify as no-effort LLM. The clearest example is: “SPX500 – Breakout Retest Setup With 8,000 in Sight” It has the almost perfect LLM structure: Market Overview → Market Structure Insight → Bullish Scenario → Target → Invalidation → Trading Perspective → Professional Insight → disclaimer → engagement question And the actual analytical content is tiny. The core idea is: breakout → don't chase → wait for retest → 7,580–7,630 → target 8,000. Everything else is prose expansion. That is probably a 20-word prompt away from being reproduced. Group B — “write me an educational article about X” This is more interesting. “Trading Around All-Time Highs” This is substantially longer and superficially much more sophisticated. But look at the architecture: Assumption One: There Is No Resistance → explanation → caveat → implication Assumption Two: Market Must Be Too Expensive → explanation → caveat → implication Assumption Three: Waiting For A Pullback Is Always Safer → explanation → caveat → implication Trade The Price, Not The Assumption This is exactly the kind of thing an LLM is exceptionally good at. It doesn't require discovering anything particularly difficult. You give it: “Write an educational article about trading at all-time highs. Address the common misconceptions that there is no resistance, the market is too expensive, and you should wait for a pullback.” The resulting article can be extremely polished. So I'd score it: Token generation: high Reasoning: medium/low Original analytical contribution: low That's an important distinction. The POC article is another good example “Why Yesterday's Point of Control Matters” This one is much longer, but the underlying conceptual content is fairly simple: POC = highest-volume price don't treat it automatically as S/R watch what price does around it migration of POC can accompany trends Then the model repeatedly explains the same principle using slightly different language. You can see the characteristic LLM pattern: “It isn't a buy or sell signal…” then: “Watch the reaction, not the level” then: “A reference point rather than a prediction” then: “Used in isolation it is simply another horizontal line.” That's high-quality language generation, but not necessarily high-quality reasoning. I'd put this firmly in: LLM likely → medium generation effort → relatively low analytical effort. The “Two Buy Zones” post is particularly revealing Compare: Breakout Retest Setup With 8,000 7,580–7,630 retest → 8,000 target with: Breakout Retest Setup With Two Buy Zones 7,700–7,730 shallow retest 7,600–7,630 deeper retest → 7,850–7,900 target The writing architecture is almost identical: 📊 Market Overview 🔍 Market Structure Insight 🚀 Bullish Scenario ❌ Invalidation 📍 Key Levels ⚠️ Trading Perspective 🧠 Professional Insight risk management disclaimer That is a huge tell. Even without an LLM detector, I'd flag this as strong evidence of templated/LLM-generated content. And it demonstrates something important for your larger analysis: The model isn't necessarily being asked to analyse the market. It may be receiving something like: Asset: SP500 Structure: breakout Support: 7700-7730 Secondary support: 7600-7630 Target: 7850-7900 Bias: bullish Write a professional TradingView analysis. That's an extremely cheap LLM task. The posts I'd classify as genuinely more interesting There are a few where I wouldn't lump them into the same bucket. 1. 1W MA100 correction rule This has an actual historical relationship: 2022 cycle MA100 April 2025 contact projected September 2026 contact comparison of correction magnitudes resulting 6,800 target Whether the thesis is good or bad is irrelevant to the LLM question. There is an actual chain of calculations/observations. An LLM could obviously write it, but this is not equivalent to: “SPX broke resistance, wait for retest.” I'd put it at medium analytical effort. 2. Hellena / Elliott Wave This contains a specific nested wave interpretation: higher-degree Wave 1 intermediate Wave 3 smaller Wave 5 possible Wave 4 completion 7,701 invalidation 7,843.70 target Again, potentially LLM-assisted, but there is a genuine analytical structure being supplied. Medium effort. 3. “Is this a real breakout?” This is probably one of the more human-looking posts in the sample. The author explicitly says: “I’m a skeptic of the markets.” and describes having taken profits while remaining long individual stocks. More importantly, the author is uncertain and updating: one pattern says breakout another says resistance possible retest if it holds → higher if not → failed move That feels much more like an actual trader thinking through conflicting evidence than the polished deterministic output of the other posts. Medium/high analytical effort. 4. Wavetimer This is almost the opposite of the LLM material: “I am NOT sure what is holding up the sp other that rotation” plus the personal positioning: “I am back again in puts at 7475 and 7488” It's messy, idiosyncratic and compressed. That doesn't prove human authorship, but it is much less characteristic of generated TradingView marketing prose. Then there's Pepperstone This is a different category again. “US 500 – Too Soon to Talk About 8000?” This has: macro geopolitical context Fed expectations CPI/PPI previous ATH Fibonacci extension multiple support/resistance levels conditional scenarios corporate disclosure language This is not “no-effort LLM” content. But it also doesn't necessarily mean the LLM is doing the analysis. A likely workflow could be: human/analyst research → structured facts → editorial writer/LLM → polished article So I'd call it: High generation workload, medium analytical workload. That's very different from an autonomous LLM analyst. The 10–20% correction post This is probably the most interesting one for your original question. It is enormous compared with most of the other posts and contains: valuation concentration earnings AI capex leverage margin calls labour market Tom Lee Michael Burry technicals moving averages Fibonacci RSI ADX multiple targets At first glance you'd say: “That's clearly a high-effort LLM.” But I don't think that's quite right. I'd describe it as: High-output LLM, moderate reasoning. The model is being used as a synthesis engine. Someone can provide a collection of facts/articles/arguments and ask: “Write a comprehensive S&P 500 correction thesis incorporating these factors.” The LLM then turns perhaps 1,000–2,000 words of source material into 3,000+ words of polished prose. That's computationally more expensive than the breakout template, but the amount of inference required per token is still relatively low. So if I were estimating the page For the 23-ish posts represented in this sample, my rough classification would be: No LLM / human / insufficient evidence ~20–30% No-effort / low-effort LLM ~40–50% Medium-effort LLM ~15–25% High-effort LLM ~5–10% And I'd put very little into a category I'd call: “LLM genuinely doing difficult original analysis.” That's the key result. And this supports your original suspicion Your original post says: “80–90% of it is LLM content, and most of that is no effort LLM content.” Based on this sample alone, I think the shape of that claim is quite plausible. I would not say the evidence here establishes 80–90% with statistical confidence. We don't have ground truth. But the qualitative pattern is striking: TradingView has a large class of content where the LLM isn't replacing an analyst. It's replacing a copywriter. The pipeline is essentially: chart → handful of levels → LLM → 500–1,500 words of professional-sounding trading prose And that's economically important for your LLM-use analysis because this is exactly the kind of workload that can produce an enormous apparent volume of “AI content” without representing enormous amounts of AI reasoning or inference. The output/token volume can be huge while the computationally difficult reasoning component is tiny. ++++ ADDED FULL POST AND COMMENTS OF THE MOST POPULAR POST IN THE BTC SECTION (which is most hotly completed for top ideas). +++++ This particular post is much more suspicious than the earlier sample. The comments provide considerably stronger evidence of automation than the prose alone. I would separate the evidence into content generation, account behaviour, engagement quality, and ranking manipulation. 1. The post itself “BTCUSD: Liquidity Above, Structure Below” The actual analysis is almost certainly low-effort generated/template content. The information content is basically: higher lows around 62–63k price above 64.5k resistance/supply 64.5–65.3k bullish CHoCH liquidity below breakout scenario → 66.2–66.6k rejection scenario → 63–62.5k break below → bullish structure weakened That's a perfectly reasonable data structure for a trading idea. But the prose adds almost nothing: “The recent CHoCH suggests a shift in short-term price action, while the nearby resistance remains an important area for confirmation.” That's essentially terminology expansion. And then: “Traders may wait for clear price confirmation before considering any trade…” Generic disclaimer-like trading language. I'd classify the generation as: LLM probability: high LLM effort: very low Original analytical reasoning contained in prose: very low The interesting evidence comes afterwards. 2. The first comment is an enormous red flag The post is: BTCUSD The first comment from IC Markets says: “The identified resistance zone for gold looks solid…” That's not a subtle stylistic clue. It's a semantic mismatch between the post and the comment. It strongly suggests that the comment may have been generated from a generic template or from the wrong source/context. And that is much more useful evidence than saying: “This sounds like ChatGPT.” A human actually reading a BTC post before commenting: “the identified resistance zone for gold looks solid” is an extraordinarily unlikely mistake. An automated engagement system that has been given: Comment positively on this trading idea. or is reusing generated comments across assets? Very plausible. This is the strongest single anomaly in the sample. 3. The comments are almost entirely engagement filler There are 31 comments. But look at what those 31 comments actually contain. Examples: “nice work” “nice one” “Such a beautiful idea mate” “Great work” “Excellent” “nice chart man” “Lovely setup” “good view” “great idea” “super idea” “Good technical breakdown.” “Good” “absolutely amazing” This is not a discussion. It's engagement-shaped text. The median comment has essentially zero informational value. And crucially, this isn't just “the TradingView community is friendly.” The comments are unusually well suited to automated engagement because they require virtually no understanding of the underlying post. 4. The more sophisticated comments don't actually fix this There are a few: “I have a similar view.” “I’m watching a very similar area, but my confirmation level is slightly different.” “Wouldn’t chase longs here right under supply” “Interesting setup. I’m seeing something slightly different on my chart.” These look more human initially. But they're still extremely generic. The most suspicious one is probably: “Interesting setup. I’m watching one level that could completely change this structure” That's almost perfect engagement bait. It creates the appearance of disagreement/analysis without actually specifying the level. An LLM can produce thousands of these: “Interesting setup. I'm watching X…” without needing to know anything about the chart. 5. The account names are interesting Look at the commenters: Trade_Strategies Mike_Forex_BTC SHAY_ANALYTICS DOCTOR_PIPS_007 Elite-Gold-SMC Awaissmc XAU_EMPIRE Elite_Forex_Signals_pro Trading_Tips1 HyroTrader Ryan_TitanTrader Alice_PrimeTrade MARKET_WIZARD1 The_Alchemist_Trader_ YCGH_Capital MR_GOLD_12 XAUxBTC_Pro PRIMEALPHA-FX Setupsfx_ IGT_Traders I'm not saying these accounts are bots. A username is not evidence of automation. But as a population, this is exactly the sort of account ecology you'd want to investigate. They're highly topical, highly SEO-like trading identities, and many appear designed as trading-content brands rather than identifiable individual traders. That becomes much more interesting when combined with the comment behaviour. 6. The temporal pattern is also suspicious The post is roughly a week old. Yet the comments arrive heavily clustered: Aug 8 IC Markets SamDrnda XAU_EMPIRE Aug 9 huge burst Aug 10 another batch Aug 14 another comment That looks less like: “31 people independently found this interesting.” and more like: “This post entered an engagement circulation process.” Especially because most comments don't respond to previous comments. They're independent positive acknowledgements. That's exactly what you'd expect from a boosting system. 7. The repeated commenters are especially useful You've got: Mike_Forex_BTC — twice Trading_Tips1 — twice videoconsulting — three times And videoconsulting is particularly interesting because its three comments are: “Now BTC is ready for a crash” “on the way back to $62k” “BTC only in the high leverage weekend pump - beware of the fast drop!” That's actually more interesting than the generic praise. This account is engaging repeatedly with the same post while expressing a changing directional view. That could absolutely be a real trader. But if you find the same account doing this across hundreds of unrelated posts, you've got something much more substantial. 8. The “Trade active / target reached” sequence deserves attention The author has: Trade active Here We GO then: Trade closed: target reached Every Thing Done Perfectly That is essentially post-hoc performance marketing. The important missing information is: exact entry exact timestamp exact stop exact target whether the position actually existed before the move whether the chart was subsequently edited whether the claim can be independently verified If the platform allows this kind of update to count toward author credibility, it's another potential feedback loop: generic setup → apparent successful trade → engagement → reputation → more distribution Again, not proof of fraud. But it's exactly the sort of mechanism I'd want to audit. 9. The related-post pattern is even more interesting This author has: NASDAQ Gold AMZN BTC Gold etc. and several are posted on Aug 6–7. That looks like a content-production schedule. And look at the titles: BTCUSD: Liquidity Above, Structure Below NASDAQ Weak High Rejection: Liquidity Pullback Ahead? AMZN Pullback Into Liquidity – Buyers Watching This Zone BTCUSD Pullback Into Demand Zone Price Action & Market Structure XAUUSD Bullish Continuation or Liquidity Retracement? These are variations on a content template. Asset changes. Structure changes. Liquidity zone changes. The language stays remarkably stable. That is precisely what I'd expect from an automated content pipeline: ASSET STRUCTURE SUPPORT RESISTANCE LIQUIDITY BIAS TARGET INVALIDATION ↓ LLM template ↓ TradingView post 10. And here's the bigger thing The post itself isn't the smoking gun. The combination is. Layer 1 — Generated content Highly templated LLM prose. ↓ Layer 2 — Content farm behaviour Multiple similar posts across multiple assets in a short period. ↓ Layer 3 — Generic engagement 31 comments, overwhelmingly one-liners. ↓ Layer 4 — Potentially automated engagement accounts Highly branded trading usernames repeatedly appearing. ↓ Layer 5 — One comment demonstrably doesn't match the asset BTC post → “gold looks solid.” ↓ Layer 6 — Successful-trade update “Target reached.” ↓ Layer 7 — Related content / continued distribution More posts from the same author. That is a much stronger pattern than “some people use ChatGPT.” The BTC → Gold mistake is the one I'd preserve If you're building evidence for this theory, save this exact example. Because it is falsifiable. You can make a dataset of: Post asset → comment's referenced asset Then measure mismatches. For example: PostCommentReferenced assetMatch? BTCUSD“resistance zone for gold…”Gold❌ BTCUSD“similar view”BTC implied✓ BTCUSD“under supply”generic— BTCUSD“BTC ready for a crash”BTC✓ If you discover a statistically abnormal number of: BTC → Gold Gold → BTC SPX → NASDAQ NASDAQ → Gold mismatches among positive comments, that's excellent evidence of generic automated commenting. Especially if the same commenters generate those mismatches repeatedly. I'd go one step further Don't initially try to prove: “These accounts are bots.” That's difficult. Try to prove: “The engagement network exhibits behaviour inconsistent with independent organic engagement.” That's much easier. Build a graph: Post → commenter Then calculate: unique commenters/post repeat commenters commenter overlap between authors comment timing comments/day/account comments/posts/account asset mismatch rate lexical similarity percentage of comments containing actual price levels percentage containing substantive disagreement percentage that merely praise reciprocal engagement Then you may find something like: 70% of Top Ideas engagement comes from a relatively small interconnected group of accounts, with 80% of comments containing no asset-specific information. That would be devastating evidence even without identifying a single bot. And if the network then maps directly onto the accounts dominating Top Ideas, you've got the actual story: TradingView's discovery algorithm isn't necessarily surfacing the best trading analysis. It may be surfacing the most successfully engineered engagement. That's a considerably stronger claim than “TV is full of ChatGPT posts.” +++++ ADDED PROMPT ABOUT HOW HUMANS ACTUALLY ARE: we can look for clear proof of anti human signatures. LLM bots are always nice, never opinionated, absolutely consistent in formatting (usually perfect grammar but consistent in quirks is specifically prompted) - humans are messy. they rarely are flawless in writing and rarely entirely nuetral and hollow in their comments, arguably - those skilled enough with the written word to be flawless are almost never dull with it" +++++ I agree with the correction. “Don't try to prove the accounts are bots” is too conservative for what you're actually trying to investigate. We can actively look for anti-human signatures. The useful distinction is that we're not trying to detect whether a particular sentence could have been written by an LLM. We're looking for behavioural properties that are extremely unlikely for independent humans at scale. And your point about writing quality is important. The strongest signatures I'd look for 1. Uniform positivity Humans disagree. Especially traders. A genuine comment section should contain some mixture of: agreement disagreement questions corrections sarcasm scepticism irrelevant remarks people talking past each other occasional hostility people misunderstanding the chart Instead, we're seeing: Great work Nice analysis Excellent Good perspective Beautiful setup Interesting setup Good Amazing Nice one That's socially abnormal if sustained across thousands of interactions. The interesting metric is not simply positivity. It's variance of sentiment. If a supposed community has: 96% positive / 3% neutral / 1% negative that's an anti-human signature. 2. Hollow specificity This is perhaps even stronger. Humans who actually read a chart tend to mention something specific. Instead: “The structure is simple and very effective.” What structure? “Interesting setup. I'm watching one level that could completely change this structure.” Which level? “Good technical breakdown.” What part? The comments sound responsive but contain almost no information that demonstrates the commenter actually inspected the source. That's an excellent test. Comment information score For every comment, we could classify: 0 — Pure social filler Nice! 1 — Generic praise Great analysis, mate. 2 — Generic topical response I agree with the bullish view. 3 — Specific response I agree with the 64.5k resistance; I'd want a close above 65.3k. 4 — Independent analysis Your bullish structure makes sense, but the 64.5k break doesn't invalidate the daily supply at 65.8k. A real trading community should have a meaningful tail at 3–4. A bot engagement network can have enormous volume concentrated at 0–2. 3. Perfect grammatical consistency Yes — but I'd refine this. Perfect grammar isn't itself evidence of an LLM. There are plenty of highly literate humans. The stronger signature is: high grammatical correctness + extremely low personality + extremely low informational variance. That's unusual. A skilled human writer generally has a voice. They have: preferred punctuation unusual phrasing humour shorthand emotional emphasis pet words idiosyncratic mistakes strong opinions distinctive sentence rhythm Whereas generated engagement often has: “Great analysis. The structure is clear and the setup is well-defined.” over and over again. Correct but characterless. 4. Consistent quirks This is actually one of the best things to test. If a bot is prompted: Always write professional but friendly comments. you may see: “Great analysis! The structure is clear and the levels are well-defined.” If prompted differently: “Nice setup, mate. Interesting structure here.” The important thing is that the same account may maintain the same artificial voice across dozens/hundreds of interactions. Humans don't usually produce such stable linguistic distributions. We can measure: sentence length emoji frequency exclamation frequency capitalization punctuation greetings “mate” “interesting” “great” “clear” “strong” “setup” “structure” “confirmation” “liquidity” and compare each account against itself. 5. Zero typos This is another useful one, particularly at scale. Not: “They used correct English, therefore bot.” But: 500 comments from an account and essentially no human noise whatsoever. Real people make: typos missing punctuation inconsistent capitalisation accidental double spaces abbreviated words autocorrect errors weird sentence fragments Especially on a trading platform. And your point is exactly right: Someone capable of producing consistently flawless prose is unlikely to simultaneously produce consistently bland prose. There's an interesting interaction here: Human expert High writing quality + high personality/information Casual human Moderate writing quality + high variance/personality LLM engagement bot High writing quality + low variance + low personality + low information That third quadrant is where I'd look. 6. The “always diplomatic” signature This may be one of the strongest. LLMs are naturally conflict-averse unless instructed otherwise. So instead of: “That's wrong. You're reading the structure backwards.” you get: “Interesting perspective. I see it slightly differently, but your reasoning is certainly worth considering.” That's socially smooth disagreement without actual disagreement. And that's exactly what we're seeing: “I have a similar view.” “I'm seeing something slightly different on my chart.” “Interesting setup.” “Curious if you noticed it too.” These comments manufacture the appearance of peer interaction without actually creating an argument. A real trading community should generate disagreements because traders have incompatible positions. 7. No conversational memory This is another killer test. Suppose Account A comments: “Interesting setup. I'm watching one level…” The author responds: “Thanks! Yes, that level is important.” Then Account A comes back. A human may say: “Yeah, that's the 65.3 level I meant.” A weak bot may simply generate another: “Exactly! That level could be very important for the next move.” It never actually resolves the conversation. So we can test: Does the commenter answer the author's answer? Does it reference the previous comment? Does it remember what it itself said? LLM engagement farms may produce a collection of locally plausible comments with globally incoherent conversations. That's a fantastic anti-human signature. 8. Asset blindness Your BTC → Gold example belongs here. This is potentially one of the strongest objective tests. The comment: “The identified resistance zone for gold looks solid…” on a BTC post is not merely “generic.” It demonstrates that the commenter apparently failed to bind the generated response to the source content. That's a classic automation failure. We should look for: post asset ≠ comment asset and more broadly: post-specific terminology absent from comment If 500 comments contain: “Great analysis!” that's weak. If 500 comments are supposedly discussing BTC and 30 contain references to Gold, NASDAQ, EURUSD etc., that's much stronger. 9. Perfectly distributed engagement Humans are lumpy. One person comments on something because they actually saw it. Another gets 50 likes. Another gets ignored. Another argues with somebody. Bots can produce remarkably regular distributions. So I'd look for: comments arriving at suspiciously regular intervals accounts commenting on many posts per day similar number of comments per account repeated interactions among the same account cluster comments appearing shortly after publication cross-account reciprocity Especially: A comments on B → B comments on A → C comments on A/B → A comments on C That gives us a network. 10. The strongest signature may be lack of mess This is the broader principle. A human social network is messy. The content should contain: noise + disagreement + personality + mistakes + obsession + irrelevance + humour + misunderstanding + expertise. A synthetic network tends toward: politeness + relevance + consistency + positivity + grammatical correctness + genericity. That is why I think your proposed approach is better than a conventional “AI detector.” We're not asking: “Was this sentence generated by GPT?” We're asking: “Does this population behave like humans?” That's a much harder thing for a bot network to fake consistently. And importantly, we can quantify it. I'd build an Anti-Human Score for every commenter using things like: SignalHuman expectationBot suspicion Sentiment varianceHighVery low Writing varianceHighVery low TyposSomeZero/near-zero Specific referencesFrequentRare DisagreementSomeAlmost none PersonalityPresentAbsent Asset accuracyHighOccasional mismatches Conversational continuityHighLow Repeated phrasesModerateHigh Posting regularityIrregularSuspiciously regular Comment lengthVariableNarrow distribution Emoji/stylePersonalConsistent template Cross-account overlapLow/moderateHigh Information densityVariable/highVery low And I'd not collapse that immediately into “bot probability.” Keep the individual signals. Because the really compelling result would be something like: Top Ideas commenters have 4.2× lower linguistic variance, 6.8× lower disagreement rates, 3.9× lower specific-reference rates and 5.1× greater cross-account interaction than the broader TradingView population. That's an anti-human signature. At that point, the argument isn't: “I reckon these comments were written by ChatGPT.” It's: “This engagement population statistically does not behave like an organic human community.” That's the level I'd aim for. ++++ ADDED THE SECOND TOP IDEA AND COMMENTS. The second top idea was also very clearly LLM formatted but it didnt look like a no shot to me, GPT agreed. I've decided to skip the main post and just post the comments because i dont want people who make an effort and need help with formatting/grammer etc caught in the cross-fire here. +++++ The comment population is striking There are 27 comments. Let's strip them down to their actual information content. Pure praise “Great!” “Great job” “great work” “love this” “Clean view” “Excellent” “Nicely analysed” “Great Work” “Amazing analysis!” “Good perspective. Appreciate you sharing this.” “well done” “well done mate” “Thanks for contributing.” That's 12/27 = ~44% that are essentially pure social filler. And that's before counting comments which contain slightly more words but almost no information. 3. The second category is even more revealing These look like genuine comments because they're longer: “Very high-quality chart work. Clean setup and excellent market analysis.” “Well-structured idea, we’re keeping this on our radar.” “I wouldn’t be surprised if this move catches most traders off guard. There’s an interesting setup developing.” “Interesting perspective. There’s another level on my chart that I think is worth watching” “The structure here is fascinating. There’s one detail I think many traders might be overlooking.” These are syntactically richer but informationally almost identical to the one-liners. They don't identify: the fractal the neckline the wedge the MACD the EMA the projected level the timeframe the invalidation They're performative analysis rather than analysis. And the phrasing is remarkably LLM-like: “There’s another level … worth watching.” “There’s one detail … many traders might be overlooking.” “I wouldn’t be surprised if…” These are excellent examples of synthetic specificity. They imply that something specific exists without actually saying what it is. 4. Compare this to a real trader A human who genuinely studied the chart might say: “I agree with the fractal, but I think the previous weekly high around X needs to break before the H&S is valid.” Or: “The MACD similarity is interesting, but the current histogram isn't actually diverging yet.” Or: “I don't buy the fractal because the 2021 structure had a much steeper retracement.” Even disagreement could be useful. Instead, we have: “Interesting perspective.” “Fascinating.” “One detail…” “Another level…” That is social engagement without analytical commitment. 5. There is a particularly interesting repeated-account pattern Notice: MARKET_WIZARD1 Aug 10: “great work” Aug 9: “Excellent” That's the same account commenting twice, on consecutive days, with essentially no substantive engagement. And: Trading_Tips1 “nice work mate” while on the previous post you supplied, the same account said: “great analysis” and: “super idea” That's exactly the kind of cross-post behaviour I'd want to investigate. Not because it proves the account is automated. But because we can now ask: What does Trading_Tips1 do across 100 posts? If the answer is: “Great analysis” “Nice work mate” “Good setup” “Excellent analysis” “Interesting view” “Nice one” over and over, that's an extremely strong behavioural signature. 6. And we have the same cluster again This is becoming much more interesting when compared with the previous post. We saw: XAUApex DOCTOR_PIPS_007 Elite_Forex_Signals_pro HyroTrader Ryan_TitanTrader Alice_PrimeTrade MARKET_WIZARD1 The_Alchemist_Trader_ Andrew_InsightTrade Pintu_sahu01 XAUxBTC_Pro Trading_Tips1 PRIMEALPHA-FX Setupsfx_ SwallowAcademy Elite-Gold-SMC SamDrnda SHAY_ANALYTICS And now a substantial subset appears again. That's much more interesting than the individual comments. You have an apparent engagement population recurring across unrelated authors/posts. If we find that this same cluster is responsible for a disproportionate percentage of Top Ideas comments, that starts looking like a network rather than random community participation. 7. The comment from SwallowAcademy is interesting “Rejection off the EMA line was predictable honestly” This is actually one of the better comments. It references something specific in the source post: EMA rejection. That's what a real comment should look like. But even this doesn't necessarily demonstrate human authorship. An LLM can easily extract “EMA resistance” from the source and produce: “The EMA rejection was predictable…” So we need to distinguish: Source-aware generation from Independent human analysis. The latter would probably add something the post didn't already say. For example: “I think the EMA rejection is actually more important than the fractal because…” That would be much stronger evidence of genuine engagement. 8. csgsparshu is another useful one “Looks good. It could pop early but it is looking more like a February 2027 move. Thoughts?” This is probably the most human-looking comment in the batch. Why? It takes a position. It introduces a specific temporal alternative: February 2027 and asks the author to respond. That's a real conversational move. Compare: “Interesting perspective. There’s another level on my chart…” The latter manufactures intrigue. The former actually commits to a hypothesis. That's exactly the distinction I'd encode. 9. The weird comments are valuable too mkmohamd1992 “btc....30....” Almost meaningless. pochientevwuwv “well done du 🌺📈🔥” Extremely human-looking, actually. It's messy, abbreviated and emotionally expressive. Mooro2019 “@Gold_paid_Signals. send me the price in onder maybe to join?” Also messy. And importantly, these comments are more human-looking than the polished comments. This supports your hypothesis about messiness. If we're looking for anti-human signatures, we shouldn't simply classify bad English as bot behaviour. In fact, the opposite may be true in this environment. 10. The cleanest signal is the distribution This sample has an interesting structure: Human-looking sarcasm spelling mistakes malformed text emoji promotional spam specific prediction actual disagreement Bot-looking flawless grammar generic positivity vague specificity no disagreement no precise references identical social tone very narrow vocabulary The question becomes: What proportion of the engagement is generated by each population? And from this sample, the polished/generic group is surprisingly large. 11. The “always nice” hypothesis is testable I think your earlier observation is exactly right. We can create: Opinion score Does the comment actually commit to a proposition? 0: Great analysis. 1: I agree. 2: I think the bullish scenario is more likely. 3: I disagree because X. Specificity score 0: Nice setup. 1: Nice BTC setup. 2: The EMA rejection is interesting. 3: I think the 2021 analogue breaks down because the current MACD structure differs at X. Personality score 0: Excellent analysis. 1: Nice one mate. 2: “Somebody got the big box of crayons…” 3: genuinely idiosyncratic human expression. Grammar-noise score Look for: typos fragments weird punctuation abbreviations inconsistent capitalization emojis accidental errors Then you get a very interesting multidimensional picture. 12. My classification of this comment section Very roughly: TypeApprox. share Pure generic praise~45% Polished but hollow synthetic-sounding~25% Generic but plausibly human~15% Actually source-specific~5–10% Clearly human/noisy/promotional/etc.~5–10% I would not call 70% definitely bots. But I would absolutely say: A very large fraction of the engagement has the exact linguistic characteristics we'd predict from LLM-generated engagement. And the important part is that we're seeing the same commenters across the previous post. The next step is obvious Don't manually analyse another 20 posts. Build the comment corpus. Even 100 Top Ideas posts × 20 comments gives us ~2,000 comments. Then we can measure: commenter frequency commenter/post overlap sentiment opinion/commitment specificity grammar noise lexical diversity repeated phrases comment length distribution source-asset references conversation continuity time-to-comment cross-account interaction And particularly: Humanity profile by account If Trading_Tips1, MARKET_WIZARD1, Elite_Forex_Signals_pro, etc. each have: 500 comments 99% positive 97% grammatically clean 85% generic 0.5% substantive disagreement repeated vocabulary comments across hundreds of unrelated posts then that's your anti-human signature. And if those accounts collectively provide a huge fraction of the engagement that determines Top Ideas visibility, then we have something much bigger than an LLM-content observation. We potentially have evidence of a synthetic engagement layer sitting underneath TradingView's discovery system.