Why AI sucks at freelance work and real-life duties: AI Eye

Mass unemployment from AI quickly suspended

AI brokers can’t full 97% of duties on Upwork to even a primary normal.

Researchers at Scale AI and the Middle for AI Security acquired six totally different AI fashions to try 240 Upwork initiatives throughout classes, together with writing, design and knowledge evaluation after which in contrast the outcomes to the actual freelancer.

The overwhelming majority of the time, the AI fashions had been unable to finish the duties efficiently, with one of the best AI mannequin, Manus, finishing simply 2.5% of duties and incomes $1,810 out of $143,991 on supply. Claude Sonnet and Grok 4 managed to complete 2.1% of the duties.

Whereas AI brokers are good at easy and outlined duties like “generate a emblem,” the analysis discovered they’re dangerous at multi-step workflows, taking any initiative or utilizing judgment.

So that they gained’t be inflicting mass unemployment for some time but.

This backs up analysis from August at MIT, which discovered that 95% of organizations had zero return on the collective $30 billion they’d invested in AI.

Why people nonetheless have the sting over AI

AIs are good at sample matching and predicting phrases. However they’re presently fairly dangerous at constructing inner fashions of the world, in keeping with WorldTest from MIT and Foundation Analysis.

For instance, people have an inner mannequin of their very own kitchen of their minds, which permits them to find out the place the knives are, how lengthy it’ll take for the pot to boil, and to plan a sequence of actions leading to a meal. However the testing confirmed that three frontier reasoning AI fashions suck at it.

The researchers created 129 duties throughout 43 interactive worlds (spot the distinction, physics puzzles, and so on). The duties required the AIs to foretell hidden features of the world, plan sequences of actions to realize a purpose, and decide when the foundations of the surroundings modified. Then they examined 517 people on the identical issues.

The researchers concluded:

“Our evaluation reveals that people obtain near-optimal scores whereas present fashions often fail.”

People carry out higher on these types of duties as a result of we intuitively perceive environments, revise beliefs within the face of latest proof, run experiments, begin from scratch and discover strategically.

And including extra compute doesn’t all the time work, serving to in solely 25 out of 43 environments.

AI will get the information unsuitable 45% of the time

Analysis from the BBC and European Broadcasting Union discovered that ChatGPT, Copilot, Gemini and Perplexity additionally suck at reporting the information, failing towards key standards, together with accuracy, sourcing, distinguishing opinion from truth, and offering context.

— 45% of AI solutions had no less than one vital subject

— 31% had been sourced incorrectly

— 20% had been simply unsuitable with hallucinated particulars and outdated data

— Gemini was by far the worst, with vital points in 76% of its responses.

AI cowl letters get the unsuitable folks employed

The first function of a canopy letter is to tell apart between low-effort purposes. Anybody spending a day writing a very good cowl letter in response to a job advert that exhibits data of the corporate is prone to be diligent and motivated.

Sadly, new analysis on Freelancer.com means that AI-generated cowl letters have utterly compromised this sign, leading to employers hiring fewer folks, and sometimes the unsuitable ones.

In comparison with the times earlier than AI, expert employees within the prime quintile for talents are being employed 19% much less typically, and dumb bums within the lowest quintile are being employed 14% extra typically.

New robotic appears like a girl

Chinese language EV producer XPeng has unveiled theXPeng Iron feminine robotic, which bears a hanging resemblance to a human. It has an analogous backbone motion to people, and pores and skin stretched over smooth 3D lattice buildings mimics the human physique.

It’s due to enter manufacturing early subsequent yr, however the firm says it requires an excessive amount of compute to be used within the dwelling, so it’ll seemingly be used for industrial purposes first, like introducing vehicles to clients at Xpeng shops.

The primary catwalk from @XPengMotors XPENG IRON feminine robotic seems to cross the uncanny valley in its strolling!
Mass manufacturing is slated to start finish of 2026, with preliminary deployments in industrial places. China appears to be within the lead in robotics advances & manufacturing. pic.twitter.com/CiSsFMkW7Y

— Derya Unutmaz, MD (@DeryaTR_) November 5, 2025

80% of ransomware assaults are made up

A brand new paper from MIT Sloan researchers and Secure Safety makes the terrifying declare that 80% of ransomware assaults are AI-driven.

Rethinking the Cybersecurity Arms Race examined 2,800 ransomware assaults and concluded that “adversarial AI is now automating total assault sequences,” creating malware, phishing campaigns and deepfake cellphone requires social engineering.

However different ransomware specialists say the statistic appears like 100% bullshit.

Researcher Kevin Beaumont tracks ransomware on-line and says generative AI isn’t a serious a part of any of them.

“The paper is nearly full nonsense. It’s jaw droppingly dangerous. It’s so dangerous it’s troublesome to know the place to start out.”

The researchers listing long-defunct ransomware, equivalent to Emotet and Conti, as “AI-powered” and incorrectly classify IBM’s DeepLocker as malware.

“The paper was so absurd I burst out laughing,” wrote researcher Marcus Hutchins.

David Sacks is anxious about Orwellian AI

Crypto and AI Czar David Sacks informed the a16z Podcast he worries that the censorship we’ve seen on social media and search engines like google in recent times will develop into totally dystopian with AI fashions.

“I virtually really feel just like the time period ‘woke AI’ is inadequate to clarify what’s occurring as a result of it one way or the other trivializes it,” he says. 

“What we’re actually speaking about is ‘Orwellian AI.’ We’re speaking about AI that lies to you, that distorts a solution, that rewrites historical past in actual time to serve a present political agenda of the people who find themselves in energy.”

“To me, that is the largest danger of AI… It’s not The Terminator, it’s 1984.”

Third time fortunate for Coke Christmas advert

Coke acquired roasted for its AI-generated Christmas industrial final yr (itself a remake of a 1995 advert), so that they remade the remake to spotlight how a lot AI video era has improved previously yr. 

“Final yr, folks criticized the craftsmanship. However this yr the craftsmanship is ten instances higher,” Pratik Thakar, international vp and head of generative AI at Coca-Cola, informed The Hollywood Reporter.

Effectively, possibly 10% higher.

The 60-second industrial was spliced collectively from two- to three-second AI-generated clips. 5 staff generated 70,000 individually generated clips, which signifies that about 2000 clips had been generated for every clip used within the advert. However it solely took a month to place collectively, in comparison with the year-long manufacturing of their live-action commercials.

A survey by Attest discovered that round 46% of shoppers within the US, UK and Australia don’t like AI-generated imagery in adverts.

Google’s Undertaking Suncatcher and different Google AI information

— There’s not sufficient electrical energy accessible for the deliberate enlargement of AI right here on Earth, so Google has give you the progressive thought of launching the information facilities into area.

The search engine big has unveiled Undertaking Suncatcher, which gives a imaginative and prescient for “fleets of satellites geared up with photo voltaic arrays” benefiting from “near-constant daylight.” 

Described as a “moonshot to someday scale machine studying in area,” it’ll launch two prototype satellites in early 2027 fitted with customized AI chips from Google’s ground-based knowledge facilities.

— Google CEO Sundar Pichai says the Gemini app now has greater than 650 million month-to-month lively customers. That’s up from round 350 million customers in March, and 90 million final October.

— The corporate was compelled to drag its Gemma AI mannequin from AI Studio after the bot claimed Senator Marsha Blackburn had pressured a state trooper for pharmaceuticals and engaged in non-consensual habits. In a letter to Pichai, Blackburn mentioned it was “not a innocent hallucination” however defamation. 

Subscribe

Essentially the most partaking reads in blockchain. Delivered as soon as a
week.

Andrew Fenton

Andrew Fenton is a journalist and editor with greater than 25 years expertise, who has been masking cryptocurrency since 2018. He spent a decade working for Information Corp Australia, first as a movie journalist with The Advertiser in Adelaide, then as Deputy Editor and leisure author in Melbourne for the nationally syndicated leisure lift-outs Hit and Switched on, revealed within the Herald-Solar, Day by day Telegraph and Courier Mail.

His work noticed him cowl the Oscars and Golden Globes and interview a few of the world’s greatest stars together with Leonardo DiCaprio, Cameron Diaz, Jackie Chan, Robin Williams, Gerard Butler, Metallica and Pearl Jam.

Previous to that he labored as a journalist with Melbourne Weekly Journal and The Melbourne Instances the place he gained FCN Finest Characteristic Story twice. His freelance work has been revealed by CNN Worldwide, Impartial Reserve, Escape and Journey.com.

He holds a level in Journalism from RMIT and a Bachelor of Letters from the College of Melbourne. His portfolio consists of ETH, BTC, VET, SNX, LINK, AAVE, UNI, AUCTION, SKY, TRAC, RUNE, ATOM, OP, NEAR, FET and he has an Infinex Patron and COIN shares.

Comply with the creator @andrewfenton