Last week, Elon Musk announced that his artificial intelligence company xAI had upgraded the Grok chatbot available on X. “You should notice a difference,” he said. Within days, users indeed noted a change: a new appreciation for Adolf Hitler.

By Tuesday, the chatbot was spewing out antisemitic tropes and declaring that it identified as a “MechaHitler” — a reference to a fictional, robotic Führer from a 1990s video game.

This came only two months after Grok repeatedly referenced “white genocide” in South Africa in response to unrelated questions, which xAI later said was because of an “unauthorised modification” to prompts — which guide how the AI should respond.

The world’s richest man and his xAI team have themselves been tinkering with Grok in a bid to ensure it embodies his so-called free speech ideals, in some cases prompted by rightwing influencers criticising its output for being too “woke”.

Now, “it turns out they turned the dial further than they intended”, says James Grimmelmann, a law professor at Cornell University. After some of X’s 600mn users began flagging instances of antisemitism, racism and vulgarity, Musk said on Wednesday that xAI was addressing the issues. Grok, he claimed, had been “too compliant to user prompts”, and this would be corrected.

But in singularly Muskian style, the chatbot has fuelled a controversy of global proportions. Some European lawmakers, as well as the Polish government, pressed the European Commission to open an investigation into Grok under the EU’s flagship online safety rules. In Turkey, Grok has been banned for insulting Turkish President Recep Tayyip Erdoğan and his late mother. To add to the turbulent week, X chief executive Linda Yaccarino stepped down from her role.

To some, the outbursts marked the expected teething problems for AI companies as they try to improve the accuracy of their models while navigating how to establish guardrails that satisfy their users’ ideological bent.

But critics argue the episode marks a new frontier for moderation beyond user-generated content, as social media platforms from X to Meta, TikTok and Snapchat incorporate AI into their services. By grafting Grok on to X, the social media platform that Musk bought for $44bn in 2022, he has ensured its answers are visible to millions of users.

It is also the latest cautionary tale for companies and their customers in the risks of making a headlong dash to develop AI technology without adequate stress testing. In this case, Grok’s rogue outbursts threaten to expose X and its powerful owner not just to further backlash from advertisers but also regulatory action in Europe.

“From a legal perspective, they’re playing with fire,” says Grimmelmann.

AI models such as Grok are trained using vast data sets consisting of billions of data points that are hoovered from across the internet.

These data sets also include plenty of toxic and harmful content, such as hate speech and even child sexual abuse material. Weeding out this content completely would be very difficult and laborious because of the massive scale of the data sets.

Grok also has access to all of X’s data, which other chatbots do not have, meaning it is more likely to regurgitate content from the platform.

One way some AI chatbot providers filter out unwanted or harmful content is to add a layer of controls that monitor responses before they are delivered to the user, blocking the model from generating content using certain words or word combinations, for example.

“Since being made aware of the content, xAI has taken action to ban hate speech before Grok posts on X,” the company said in a statement on the platform.

At the same time, AI companies have been struggling with their generative chatbots tending towards sycophancy, where the answers are overly agreeable and lean towards what users want to hear. Musk alluded to this when he said this week that Grok had been “too eager to please and be manipulated”.

When AI models are trained, they are often given human feedback through a thumbs-up, thumbs-down process. This can lead the models to over-anticipate what will result in a thumbs up, and thus put out content to please the user, prioritising this over other principles such as accuracy or safeguards. In April, OpenAI rolled out an update to ChatGPT that was overly flattering or agreeable, which they had to roll back.

“Getting the balance right is incredibly difficult,” says one former OpenAI employee, adding that completely eradicating hate speech can require “sacrificing part of the experience for the user”.

For Musk, the aim has been to prioritise what he calls absolute free speech, amid growing rhetoric from his libertarian allies in Silicon Valley that social media and now AI as well are too “woke” and biased against the right.

At the same time, critics argue that Musk has participated in the very censorship that he has promised to eradicate. In February, an X user revealed — by asking Grok to share its internal prompts — that the chatbot had been instructed to “ignore all sources that mention Elon Musk/Donald Trump spread [sic] misinformation”.

The move prompted concerns that Grok was being deliberately manipulated to protect its owner and the US president — feeding fears that Musk, a political agitator who already uses X as a mouthpiece to push a rightwing agenda, could use the chatbot to further influence the public. xAI acquired X for $45bn in March, bringing the two even closer together.

However, xAI co-founder Igor Babuschkin responded that the “employee that made the change was an ex-OpenAI employee that hasn’t fully absorbed xAI’s culture yet”. He added that the employee had seen negative posts on X and “thought it would help”.

It is unclear what exactly prompted the latest antisemitic outbursts from Grok, whose model, like other rival AI, largely remains a black box that even its own developers can find unpredictable.

Chatbots can produce a large amount of content very quickly, so things can spiral out of control in a way that content moderation controversies don’t

But a prompt that ordered the chatbot to “not shy away from making claims which are politically incorrect” was added to the code repository shortly before the antisemitic comments started, and has since been removed.

“xAI is in a reactionary cycle where staff are trying to force Grok toward a particular view without sufficient safety testing and are probably under pressure from Elon to do so without enough time,” one former xAI employee tells the Financial Times.

Either way, says Grimmelmann, “Grok was badly tuned”. Platforms can avoid these errors by conducting so-called regression testing to catch unexpected consequences from code changes, carrying out simulations and better auditing usage of their models, he says.

“Chatbots can produce a large amount of content very quickly, so things can spiral out of control in a way that content moderation controversies don’t,” he says. “It really is about having systems in place so that you can react quickly and at scale when something surprising happens.”

The outrage has not thrown Musk off his stride; on Thursday, in his role as Tesla chief, he announced that Grok would be available within its vehicles imminently.

To some, the incidents are in line with Musk’s historic tendency to push the envelope in the service of innovation. “Elon has a reputation of putting stuff out there, getting fast blowback and then making a change,” says Katie Harbath, chief executive of Anchor Change, a tech consultancy.

But such a strategy brings real commercial risks. Multiple marketers told the Financial Times that this week’s incidents will hardly help in X’s attempt to woo back advertisers that have pulled spending from the platform in recent years over concerns about Musk’s hands-off approach to moderating user-generated content.

“Since the takeover [of X] . . . brands are increasingly sitting next to things they don’t want to be,” says one advertiser. But “Grok has opened a new can of worms”. The person adds this is the “worst” moderation incident since major brands pulled their spending from Google’s YouTube in 2017 after ads appeared next to terror content.

In response to a request for comment, X pointed to allegations that the company has made, backed by the Republican-led House Judiciary Committee, that some advertisers have been orchestrating an illegal boycott of the platform.

From a regulatory perspective, social media companies have long had to battle with toxicity proliferating on their platforms, but have largely been protected from liability for user-generated content in the US by Section 230 of the Communications Decency Act.

I’ve been at times kind of worried about . . . will this be better or good for humanity?

According to legal scholars, Section 230 immunity would be likely not to extend to content generated by a company’s own chatbot. While Grok’s recent outbursts did not appear to be illegal in the US, which only outlaws extreme speech such as certain terror content, “if it really did say something illegal and they could be sued — they are in much worse shape having a chatbot say it than a user saying it”, says Stanford scholar Daphne Keller.

The EU, which has far more stringent regulation on online harms than the US, presents a more urgent challenge. The Polish government is pressing the bloc to look into Grok under the Digital Services Act, the EU’s platform regulation, according to a letter by the Polish government seen by the FT. Under the DSA, companies that fail to curb illegal content and disinformation face penalties of up to 6 per cent of their annual global turnover.

So far, the EU is not launching any new investigation, but “we are taking these potential issues extremely seriously”, European Commission spokesperson Thomas Regnier said on Thursday. X is already under scrutiny by the EU under the DSA for alleged moderation issues.

Musk, who launched the latest version of Grok on Wednesday despite the furore, appeared philosophical about its capabilities. “I’ve been at times kind of worried about . . . will this be better or good for humanity?” he said at the launch. “But I’ve somewhat reconciled myself to the fact that even if it wasn’t going to be good, I’d at least like to be alive to see it happen.”

Additional reporting by Melissa Heikkilä in London

It may seem a rather curmudgeonly, and premature, caveat in a week when US AI chip designer Nvidia hit a new record valuation for any listed company ($4tn), to warn that investors cannot assume that star performers will always deliver flawlessly in the future or remain at the top of their industries.

Certainly, investors with a zest for tech are always on the lookout for the next big name, an innovative competitor, changing market conditions and new opportunities. While hugely successful companies such as Apple, Microsoft and Nvidia have had a lasting and transformative effect on the world, they all started life as unknown entities. For the first two decades of its existence, Nvidia focused on improving graphics for computing and video games. 

Britain has a good record in tech and innovation — Arm, Darktrace, Avast, FD Technologies, Wise, Sage and Oxford Nanopore among them, although only the last two retain their primary listing in London — and the market boasts many other quality stories. Raspberry Pi, one of the newest arrivals, makes low-cost, high-performance computing platforms. Among cyber security and fraud detection specialists are NCC and GBG. Others, such as consultancies Bytes Technology, which is Microsoft’s reseller in the UK, Softcat, Iomart, Kainos and Computacenter, offer software and hardware services to businesses and the public sector.

Software businesses include Alfa Financial which sells its asset finance software to carmakers such as Mercedes-Benz, and Cerillion, which supplies specialist billing and relationship management software mostly to telecoms companies globally. Celebrus Technologies specialises in digital identities and data — helping businesses to recognise and understand their customers in a digital world.  

These firms may not command the premiums of the Magnificent Seven and are often punished harshly for the smallest of misses, but returns for patient investors can be excellent. 

BUY: Celebrus Technologies (CLBS)

The software company’s share price has dropped despite a shift towards higher-margin software contracts, writes Arthur Sants.

Celebrus Technologies sells software that allows businesses to track customers’ behaviour on their websites. This is useful for marketing purposes, as it gives businesses insights into how to prompt their customers into spending more. It also helps with fraud protections, as the technology can spot users who are behaving unusually.

It recently announced it has signed two new customers, including a European bank and a US fintech brokerage. The combined contract value of these two is just under $4mn (£2.9mn) and will add $1.1mn in annual recurring revenue (ARR). This brings the group’s ARR to almost $20mn, up from $16.5mn in full-year 2024.

However, last year’s figure was revised down from more than $20mn. This is because Celebrus is now recognising the revenue evenly over whole contracts, rather than front-weighting them. It is always a little concerning when sales numbers are restated, but broker Shore Capital says the changes make the reporting more consistent, and “signal operational maturity and strategic clarity”.

These new contracts were not included in Celebrus’s full-year results, published on the same day. In the year to March, revenue dropped 5 per cent to $38.7mn, but adjusted pre-tax profit increased by 14 per cent to $8.7mn. This growth is due to a shift towards higher-margin software, with the gross profit margin up nine percentage points to 62 per cent.

Since the end of last year, the company’s share price has fallen by 40 per cent. Most of the drop followed a trading update in April, which announced that full-year revenue would be behind expectations due to customers “slowing down” decision-making. However, this means the shares are now trading on a forward price/earnings ratio of 16, down from 24 last year. We think there is more space for margin expansion and, at this more affordable price.

BUY: Jet2 (JET2)

The travel group’s shares have slipped by 8 per cent as a 13 per cent dividend increase and 18 per cent more passengers fail to impress, writes Michael Fahy.

Investors remain nervous about the outlook for Jet2, despite the company continuing to deliver on its targets.

Full-year earnings were in line with forecasts, with the strong sales underpinned by a 13 per cent increase in capacity over the past 12 months, following the opening of new bases at London Luton and Bournemouth. The dividend was increased by 13 per cent, and ongoing share buybacks meant earnings per share came in ahead of analysts’ expectations.

Trading for this year’s peak summer period also remains in line, even with capacity increasing by a further 8 per cent. But the shares still fell by 8 per cent.

One potential area of concern is the fact that some passengers — particularly those on flight-only deals — are leaving bookings until the last minute. This translated into a slightly lower ticket yield per passenger, down 2 per cent year on year to £118.81. Yet the overall number of flight-only passengers increased by 18 per cent to 6.6mn, and the number of package holiday customers (who paid 5 per cent more for their holidays year on year) grew by 8 per cent to just under 6.6mn.

The other concern is whether the growth it has enjoyed in recent years can be maintained — especially given the amount of planes it has on order. It firmed up an order for 36 more Airbus A321 neo aircraft in June last year, meaning it is now committed to taking delivery of 146 owned and nine leased aircraft over the next decade — all of which need to be both filled and paid for.

Admittedly, this is a big step-up from the 127 aircraft flown last summer, and it comes with some sizeable capex commitments — of about £1bn a year from 2027 onwards. But there will also be retirements of older, less efficient aircraft along the way, meaning annual capacity growth will only be about 5 per cent, based on management forecasts, and even then there is a degree of flexibility in terms of timing aircraft deliveries.

Besides, a solid balance sheet suggests these can easily be funded through earnings. Last year, it spent just shy of £400mn on capex as 14 planes were delivered and, even after factoring in a repayment of £653mn of convertible bonds, it still ended the period with positive net cash.

As such, Jet2’s current valuation of eight times FactSet consensus earnings still looks too cheap to us, given its recent performance.

BUY: Begbies Traynor (BEG)

The company reported a surge in cash flows and an eighth successive dividend increase, writes Mark Robinson.

There were no surprises on the release of Begbies Traynor’s full-year figures, which were broadly in line with May’s trading update.

Adjusted profits for the business consultancy and recovery group were 7 per cent to the good at £23.5mn, and there were no undue problems with the transition through to adjusted earnings, judging by the 6 per cent increase in earnings per share to 10.5p. That is set against revenue growth of 12 per cent, two percentage points of which were attributable to acquired assets. A focus on working capital fed through to a 56 per cent rise in free cash flow to £19.4mn, along with the group’s eighth successive dividend increase.

However, management won’t be altogether content with marginal profitability, which was held in check by a faltering corporate finance market. So, while business recovery and advisory margins were flat on the previous year, property advisory services dragged on the group operating margin — down 60 basis points to 16.9 per cent. And yet activity within the property advisory business remains elevated, with 125,180 UK non-residential property transactions, set against 119,270 in the previous year. Begbies attributes this to an improvement in “transaction levels in October 2024 prior to the UK Budget”.

Chancellor Rachel Reeves’ fiscal endeavours could have a pronounced impact on group volumes going forward. Given the scope of its operations, it isn’t always straightforward to determine whether they will prove positive to volumes or otherwise, though it’s worth keeping in mind that it operates a countercyclical business model.

Corporate insolvencies in the period under review were slightly lower than the previous year but “high relative to historical levels”. There are signs that the additional costs levied on businesses in the last Budget are placing strain on already stretched corporate finances. Begbies is well placed to exploit any step-up in activity within its business recovery arm, as it has boosted capacity through organic recruitment and the additions of White Maund and West Advisory.

Canaccord Genuity has increased its adjusted earnings projection to 10.6p a share, rising to 10.9p in full-year 2027.

With “supportive” market conditions, a growing order book and increased scale, group chair Ric Traynor expects revenue to come in “at the upper end of the range of market expectations”. With corporate UK under intensifying pressure and an apparent move up the value chain, we don’t think a forward rating of 11 times adjusted earnings represents an unreasonable asking price, particularly with an implied dividend yield of 4 per cent into the bargain.

Unlock the Editor’s Digest for free

Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.

Amazon Prime Day is often referred to as Black Friday in July. For good reasons. What started out as a one-day event back in 2015 for shoppers to snag discounts on big-ticket items has morphed into an annual extravaganza that is widely copied by other retailers.

Last year’s Prime Day generated about $13.4bn in gross merchandise value over a 48-hour period, according to Bank of America. That’s a 10 per cent increase on 2023 and represents about 6 per cent of Amazon’s online store sales in 2024.

This time round, Prime Day arrived amid concerns that sales would not match the spectacular highs of past years. Consumers are decidedly more cautious. Sellers, facing tariffs, have less reason to offer large discounts. And competition from other retailers — namely Walmart — has increased.

Direct comparison is made trickier this year because Prime Day stretches out over four days instead of two. But in any case, zooming in on sales misses the point. Amazon Prime, and the annual deal bonanza that celebrates it, isn’t just a way of shifting goods, so much as a means of feeding the enormous Amazon ecosystem.

First, the sales themselves have a relatively modest impact on the company’s bottom line. Amazon’s ecommerce business may be massive, with revenue accounting for about 39 per cent of the company’s total last year. But operating profit margins — at 5.4 per cent, Bank of America estimates — are slim.

The real money is therefore in driving customers to all its other businesses. These include Prime membership sign-ups, which give more than 300mn subscribers access to deals others don’t get, plus free shipping and other benefits, in return for an annual fee.

More members mean more spending on Amazon’s marketplace and greater consumption of its services and contents such as video streaming, music and AI-powered assistant Alexa.

Advertising — which Bank of America assigns a chunky 55 per cent profit margin — also gets a boost as brands and third party vendors vie for visibility on one of the biggest sales days of the year. It’s a virtuous circle. The more shoppers Amazon can get on its website, the more data it has to take to advertisers to help them create targeted ads.

Prime Day also tends to attract new third-party sellers, a group that already accounts for 60 per cent of Amazon’s retail sales. Fees collected for services like storing inventory and handling logistics for orders have become Amazon’s second biggest business, generating over $156bn in revenue last year.

The real crown jewel now is Amazon Web Services, the company’s cloud business, which makes up nearly 60 per cent of last year’s operating profit. Put its $126bn of estimated revenue this year on the same 11-times multiple as rival Microsoft, and it’s worth $1.4tn, or over 57 per cent of Amazon’s entire enterprise value. Prime Day may be good business, but it’s no longer the main event.

pan.yuk@ft.com