After a year of artificial intelligence [AI] errors, it’s clear that the robot revolution is not as perfect as claimed.

By: Andrew Orlowski / The Daily Telegraph

It has been two years since the launch of ChatGPT-3, and we’ve learned a lot during that time.

We now know, for example, that Hillary Clinton was the first woman president of the United States, thereby correcting a historical injustice. Generative artificial intelligence has offered us pasta recipes that include gasoline as an ingredient and advised us that, for the sake of our health, we should remember to eat a pebble every day.

My favorite is this: astronauts “have met cats on the Moon, played with them and taken care of them” – according to Google. Apparently, this is what Neil Armstrong was talking about when he said that the “small step” was that of a cat. How foolish we were not to notice what, looking back, was a fairly clear hint from NASA about a feline takeover of space.

History, as it turns out, is full of other surprises. The Vikings who invaded medieval Europe and, likewise, the Nazi army in the Second World War had impressive diversity policies, with many members being of African-American or Native American descent – according to Google.

Artificial intelligence has tried to improve educational standards. It insists that the word “strawberry” has two “r”s, and don’t let anyone tell you otherwise.

I’m sometimes asked about my 2013 book, The Goolag: How Google’s Empire Is Changing the World, which ChatGPT includes in summaries of my career and describes as an “outstanding work”. I’m sorry to inform you that no such book exists, nor anything similar – not by me, nor by anyone else. It is one of many works falsely attributed to me.

Now Apple has entered the game, bringing Apple Intelligence to its iPhones. It can’t get any worse, can it? It’s Apple. Unfortunately, it has already told us that Israeli Prime Minister Benjamin Netanyahu has been arrested for war crimes (wrong) and that the suspected assassin of a prominent American healthcare industry executive died by suicide in his cell (wrong). Both stories were attributed to the BBC, which was not at all happy.

Google was forced to apologize after its chatbot created a Black Nazi and a Viking as a Native American.

However, lucky Americans will soon be able to call a toll-free number (it’s 1-800-CHATGPT) and listen to content of this quality being read aloud.

Some people pay a lot to be insulted over the phone – London’s public phone boxes were once full of brochures promoting such services – so perhaps this isn’t new. But it is a clear reminder that the statistical predictions underlying Large Language Models (LLMs) often lead to terrible results, called “hallucinations”.

For this, I don’t entirely blame the artificial intelligence companies. Exaggerating and overselling a technology is simply what they do today. They take no more responsibility for the consequences than the scorpion in the fable, which asked to cross the river on a frog, promised not to sting it, but did so anyway, drowning them both.

But unlike in the fable, the scorpions of Silicon Valley always seem to make it to the other side, where they simply start all over again. If there’s a problem, you can be sure they won’t be there to help clean up the mess.

We who insist on using technology where it is unsuitable are just as responsible. Reason has been lost.

For example, Tony Blair is urging us to spend billions on AI to improve the efficiency of the National Health Service [NHS] – automating, rather than replacing, the many layers of bureaucracy.

Dominic Cummings now indulges in wild fantasies about teams of chatbots replacing the real voters who currently participate in focus groups. He even imagines these hallucinations being fed directly into another AI to produce campaign material. Cummings has spent so much time in the company of Silicon Valley snake-oil salesmen [purveyors of fake remedies] that he seems to think everyone else is stupid.

When artificial intelligence is used sensibly, it can be genuinely very useful. Machine learning has proved to be a valuable tool in data analysis, finding patterns in vast masses of data. Where errors are inconsequential, as in many signal and image-processing applications, AI has been successfully deployed. Text research assistants aren’t ready yet, but are expected to improve.

Five cases when AI went wrong:
ChatGPT speaks “Spanglish”
1. In February, OpenAI’s ChatGPT seemed to malfunction, giving nonsensical answers to questions, reverting to endless lists and speaking in “Spanglish”.
Microsoft’s Bing went out of control
2. Microsoft’s early attempt to add a chatbot to its Bing search engine ended badly. Users managed to “break” the bot, which began calling itself Sydney, declaring its everlasting love for a journalist and even asking him to get a divorce.
Meta’s racist chatbot
3. In 2022, Meta released a chatbot called BlenderBot, designed to have natural conversations, but it went out of control, spreading racist conspiracy theories, such as the claim that Jews “control the economy”.
Google’s LaMDA becomes “conscious”
4. An internal chatbot built by Google caused embarrassment after an engineer published claims that the bot had become sentient. The engineer was later fired.
“Will Smith eating spaghetti”
5. A viral video demonstrated how poor some early AI video generations were. A well-known clip from 2023, created with a tool called ModelScope, showed a version of Will Smith eating spaghetti – while his face was grotesquely distorted.

The problem for the technology industry is that you can’t create a huge stock-market bubble or make the political class hang on your every word if you’re promoting a software program that is “actually sometimes quite useful”. It isn’t revolutionary enough, and that has been the only way they know to sell their products.

Our credulity has been questioned for decades. In the 1960s, a computer scientist named Joseph Weizenbaum developed two highly successful conversational programs, ELIZA and DOCTOR, and users’ reactions astonished him. People talked for hours, revealing their secrets to the software. It was their excessive credulity, not the technology, that turned Weizenbaum from a supporter of artificial intelligence into a technology skeptic.

Simon Willison, a widely followed British developer in Silicon Valley, reminds us that “LLMs are useful tools for thinking, but terrible tools for delegating decision-making. Every time someone hands actual decision-making authority to a vague random-number generator, it’s a recipe for disaster.”

Many people aren’t paying attention to the experts. In early 2024, the respected Institute for Public Policy Research (IPPR) published a report that generated a predictable headline for The Guardian: “‘AI apocalypse’ could eliminate eight million UK jobs, report says”. Last year, Goldman Sachs said that 300 million jobs – worldwide – could disappear.

But in reality, as we’ve seen, you’d have to be crazy to leave an AI system to make a final decision. If AI takes your job this year, you’ll probably be hired again very soon when your boss realizes the mistake they made. If your boss has also been replaced by an AI, he or she will be hired first.

We used to be more careful. An extraordinary presentation, created at IBM in the late 1970s, was recently uncovered. It was a set of rules software developers had to follow. “A computer can never be held responsible. Therefore, a computer should never make a management decision,” was one piece of advice. Other guidelines were just as emphatic: “AUTHORITY WITHOUT RESPONSIBILITY IS AN ELEGANT FORM OF SUICIDE,” was another instruction.

It is a message from another era, from a less cynical but more responsible generation of technologists – a generation more concerned with human well-being than getting rich quickly. /Telegrafi/