AI health check: the patient is still very sick
Follow our ongoing analysis of AI at work, this time on cryptocurrencies.
Peter Hulm·
AI does not mean artificial intelligence, of course, in spite of what its propagandists say – unless they mean surreptitiously gathered information that is not true.
What AI pretends to offer is machine learning using, usually, webwide information for its training.
So how does it do? My previous experiences were not inspiring on general news generation (LINK) or summaries (LINK).
So I tried applying AI to cryptocurrency news. That should involve simple mathematical figures and assessment. Surely this time it would do better?
Sorry.
You expected that, of course. The promoters promise major achievements. Specialists often warn of horrors. Users, you and I, for example, are disappointed. Here’s my experience.
Disaster 101
Take this example, using AITOPIA on Microsoft Edge from 22 November 2025. The headline from U.Today read: XRP's Key Metric Surges 33.9% but Price Plunges Hard, What's Next? (LINK)
AI summary
I asked AITOPIA t to summarize the article. Its version:
“XRP's circulating supply has decreased significantly, showing a 339% surge in a key metric, indicating rising institutional adoption. Despite this positive on-chain activity and growing interest due to ETF discussions, XRP's price has experienced a sharp decline, remaining at low levels.”
Judgement
339%? Can’t AI understand numbers?
What was the key metric? Nothing recorded.
In fact it was the amount of XRP burned as fees. U-Today reported: “According to data showcased by XRPSCAN, the amount of XRP burned as fees has seen a decent increase of 33.9% as of Nov. 21, printing a bullish outlook amid the growing uncertainty.”
And why didn’t it pick up on what the article predicted would come next? That was its selling point, surely?
What U-Today recorded was “the sharp resurgence in the metric despite XRP’s price decline has been attributable to growing demand from institutional investors amid the major ETF launch that happened this week. Usually, increases in the rate of burned XRP like this are often a clear indication of increased network-wide transaction activity, suggesting that more users are actively utilizing the blockchain.”
Verdict: AI added nothing and omitted the essentials.
The next article summary did fine, except that it reported the essential point for the story’s optimism on SOL’s rebound – daily active addresses – reported the month’s increase in active addresses as 58% not the actual 5.8%. Which meant I had to check the story against the summary and correct the AI product.
The corrected version: “Despite a 49% price decline from its September high, Solana's network activity is experiencing a significant surge, indicated by a 5.8% increase in daily active addresses, perhaps leading to rebound.”
Verdict: don’t trust the math.
What next for AI?
The What next? title is common in crypto reporting. I obtained a summary for a cryptonews analysis of XRP’s future in view of its dropping price: this did comparatively well compared to previous results but offered no more than a bland recommendation at the end: “Overall, the analysis highlights [...] the importance of watching the 1.80 level for future price movements.”
This is what cryptonews said: “XRP price prediction as $2 support breaks and the $1.80 zone becomes critical. Will oversold signals spark a reversal or confirm deeper weakness ahead?”
That told you all you need to know, and concisely.
Verdict: AITOPIA didn’t get the point and needs to learn why factors are included.
These are not small niggles. They can seriously mislead AI users. Even the last analysis can lead you astray if you don’t understand why the $1.80 zone is important: you need to know that you should check whether the price indicates XRP is oversold or overblown.
Major news, no thank you
My fourth test article had this news headline on nusereal.com: Nigeria sees one of worst mass abductions as 315 taken from Roman Catholic school.
AITOPIA’s summary ignored the identification of the school as Roman Catholic and once again got its figures wrong. It said: “The Christian Association of Nigeria reported that 303 students and 12 teachers were taken by armed gunmen in the early hours of Friday morning, a figure that was revised upwards after further verification.”
No. It was revised upwards from 215 pupils not 303.
The original BBC report (LINK) justified its headline with this backgrounder: “The revised number of people taken surpasses the 276 abducted during the infamous Chibok mass abduction of 2014.”
The original article also included what I consider a major aspect: “Authorities in Niger state said the school had disregarded an order to close all boarding facilities following intelligence warnings of a heightened risk of attacks. The school has not commented on that claim.”
AITOPIA didn’t think this was worth including.
Verdict: AITOPIA really screwed this one up.
In all, this behaviour made it really impossible for me to use AI as a summarizer. It didn’t save me time. In fact, I spent more than twice the time on items, having to read the stories and the summaries, and then edit the rather dull texts AITOPIA produced.