Apple wants AI to be more personal and write in your style, research paper reveals

If you have ever asked an AI to write an email or a message, you have likely noticed a familiar flaw: the results often sound too generic. Even with detailed instructions, large language models like ChatGPT or Gemini often struggle to reflect a person's actual tone or writing style. Apple now believes it has found a solution. In a new research paper to be presented at the International Conference on Machine Learning (ICML 2025), Apple researchers introduced a technique called PROSE, which is short for Preference Reasoning by Observing and Synthesising Examples. It is designed to help AI systems learn directly from a user's writing history, so future outputs can better match the user's natural tone and preferences.advertisementWhat is PROSE and how does it work?Unlike older methods such as prompt engineering or reinforcement learning from human feedback, PROSE builds a unique and evolving profile of how a person writes. It works in two key steps:1. Iterative refinement: The AI compares its generated drafts with actual user-written examples and adjusts its internal style description until the output closely resembles the user's tone.
2. Consistency verification: The system then checks whether the identified writing preferences (like 'use short sentences' or 'be friendly and casual') appear consistently across multiple samples. This ensures that the AI doesn't overly rely on just one piece of writing.The result is a self-learning system that can tailor its future writing based on your overall writing behaviour, not just a single document or prompt.While Apple didn't name any of its products in the research paper, the potential applications are clear. With the company's growing push into Apple Intelligence – its next-gen AI assistant features – tools like PROSE could allow apps like Mail or Notes to generate texts that sound much more like you.advertisementApple's new Foundation Models framework, announced recently, will let developers access Apple's local language models. That means PROSE-style personalisation could eventually power writing tools inside a wide range of third-party apps as well.Alongside PROSE, Apple also introduced a new dataset called PLUME – short for Preference Learning from User Emails and Memos. It replaces an earlier dataset, PRELUDE, and aims to better test how well AI systems can understand and replicate writing preferences.When tested using PLUME, PROSE beat existing personalisation techniques. It outperformed a similar system named CIPHER by 33 per cent and even showed better results than standard in-context learning (ICL) approaches, particularly when used with advanced models like GPT-4o. Interestingly, combining PROSE with ICL delivered the best overall results – with up to a 9 per cent improvement over ICL alone.Why personalised AI like PROSE matter Apple's PROSE research isn't happening in isolation, it follows another critical finding from the company's AI team that raises serious questions about how reliable today's large language models (LLMs) really are, especially when faced with complex tasks.In a separate paper titled The Illusion of Thinking, which was released earlier this month, Apple researchers argue that even top-tier models like ChatGPT o3, Claude 3.7 Sonnet, and DeepSeek-R1 may not be as intelligent as they seem. These models are often described as 'Large Reasoning Models' (LRMs) because they attempt to break down complicated tasks using a step-by-step 'chain of thought.' But Apple's study suggests this ability might be more fragile than expected. To test this, the researchers created custom environments — including puzzles like the Tower of Hanoi, River Crossing, and Blocks World — designed to gradually increase the complexity of the problems. The aim wasn't just to test whether the AI could get the right answer, but whether it could reason its way there logically.advertisementThe results were surprising. When tasks were simple, traditional models without complex reasoning strategies often performed better. At a medium level of complexity, reasoning-based models did briefly take the lead. But as complexity increased further, all models -- including the most advanced -- experienced what the researchers call a 'performance collapse.' In short, the AI systems that seemed smart at first struggled or gave up entirely when the problems got too hard -- despite having enough computing power to continue.This highlights why Apple's PROSE approach is timely. If AI can't truly 'think' through tough challenges yet, then at the very least, it should be able to sound more human -- by learning how you write.

Try Our AI Features

Explore what Daily8 AI can do for you:

Learn with AIFact CheckingText to SpeechAI Summary

Related Articles

Mint

2 hours ago

• Mint

The $1,999 Liberty Phone is made in America. Its creator explains how.

Next Story Ben Raab , The Wall Street Journal Purism's Liberty Phone, priced at $1,999, is the most advanced U.S.-made smartphone but lacks competitive specs compared to the iPhone. Limited by domestic supply chains and higher costs, it sources components globally. Despite scaling potential, Purism sells fewer than 100,000 units, as many parts still come from Asia. The phone's unique features cater to security-conscious users. Purism's $1,999 Liberty Phone is assembled in the U.S. with parts from the U.S. and elsewhere, including China. Photo: Purism Gift this article It is possible to build a smartphone in the U.S. right now. But it won't be as sleek or as powerful as an iPhone, and it will cost a lot more. It is possible to build a smartphone in the U.S. right now. But it won't be as sleek or as powerful as an iPhone, and it will cost a lot more. Todd Weaver's company, Purism, developed the Liberty Phone, the closest anyone has gotten. It has specs that would have been more impressive a decade ago, and it costs $1,999. President Trump has threatened steep tariffs on foreign-made smartphones to pressure companies like Apple to shift manufacturing stateside. Meanwhile, the Trump Organization is promoting a 'Made in the U.S.A." phone for $499 with specs that deem it unlikely to be built here anytime soon. Supply-chain analysts agree it's impossible to match Asia's production quality and scale for now. But Weaver's Liberty Phone, not the Trump phone, offers a unique look at the realities of domestic manufacturing. And why nobody else is doing it. The Liberty Phone's motherboard is built in-house, the chip comes from Texas, and the assembly is done at Purism's facility in Carlsbad, Calif. But not all of its parts are U.S. made: Other components come from China and other Asian countries. 'I've been working on this for 10 years and we've done everything we possibly can to build from U.S. manufacturing," Weaver says. 'There are just some parts that don't yet have a supply chain. We're gonna keep incrementing there until we can get to that point." Weaver says he can produce Liberty Phones at a rate of about 10,000 a month, but so far, he's sold fewer than 100,000. By comparison, Apple shipped around 225 million phones in 2024, according to market analyst firm Canalys. The Liberty Phone also doesn't run on Android or iOS. Its processor, produced by Dutch semiconductor firm NXP in Austin, Texas, is designed for cars, not smartphones. It runs on Purism's own PureOS, which is limited to calling, texting and web browsing, plus some basic apps like a calculator. Purism founder Todd Weaver holds up the Liberty Phone's motherboard, manufactured at the company's California facility. Photo: Purism The screen and battery come from China and the rear-facing camera comes from South Korea. Weaver says a fully U.S.-made phone is limited by a lack of domestic infrastructure. There are no companies mass-producing smartphone screens in the U.S., for example. Complications like this, even at Purism's small scale, help explain why Apple and others haven't made a serious attempt at producing premium smartphones in the U.S. 'Even if the specs were less impressive, it would take many, many years to be fully sourced out of the U.S. and not practical," says Jeff Fieldhack, a research director at Counterpoint Research. 'Cost aside, we don't have factories here building application processors, high-end displays or most of the other things in your smartphone." Weaver says the Liberty Phone costs about $650 to make. The iPhone 16 Pro Max, a much more powerful device, was estimated to cost around $550 to make in China last fall, according to TechInsights. Purism's higher U.S. labor costs are partially offset by cheaper, lower-quality parts: a basic camera, low-resolution screen and half the RAM. Weaver says the Liberty Phone isn't built to compete with an iPhone right now. He says the $1,999 retail price reflects a securely sourced phone with a vetted supply chain. About half of Purism's customers are government agencies across the U.S., he says. 'On the consumer side, it's security geeks, parents who want a phone for their kid, elderly people or people who want to avoid big tech," says Weaver. 'Someone who needs a wicked-strong camera is not our audience." Weaver estimates he could scale to building 100,000 phones a month within six months. But getting there would require investment to cover new machines, more line operators and added floor space. He has no traditional venture-capital funding, relying only on revenue and crowdfunding. Tariffs likely won't affect his costs at his current production rates because he ordered a large batch of parts when he started and still has them in supply at his facility. Weaver says that a long-term tariff on imported electronics could make the Liberty Phone's manufacturing cost more competitive, since the cheap components would only see marginal increases, and more components are likely to be built in the U.S. soon. While companies like Intel already manufacture chips in the U.S. and firms like TSMC and Micron are building domestic facilities, Fieldhack says those efforts are a small fraction of global production. Companies lack incentives to move more to American soil. 'A lot of it isn't leading edge, the cost is still high, and it would still take a long time," he says. Write to Ben Raab at Topics You May Be Interested In Catch all the Business News , Corporate news , Breaking News Events and Latest News Updates on Live Mint. Download The Mint News App to get Daily Market Updates.

Hindustan Times

4 hours ago

• Hindustan Times

Tech firms, content industry debate AI, copyright at ministry of commerce event

Who owns the data that fuels artificial intelligence (AI)? That was the central — and contentious — question debated by representatives from big tech firms and the content industry during a two-day stakeholder consultation organised by ministry of commerce and industry's department for promotion of industry and internal trade (DPIIT). A nine-member DPIIT committee will soon publish a working paper outlining whether India's copyright law needs to be updated to address challenges posed by AI. (Getty Images/iStockphoto/ Representative photo) The meetings were chaired by DPIIT additional secretary Himani Pande on June 19 and 20. At the centre of the discussion was whether tech companies should be allowed to freely mine the internet, which includes copyrighted books, articles, music, images, and videos, to train their AI models. The content industry raised concerns over their copyrighted data being used to train AI models without permission, while tech companies argued that training their models requires massive amounts of data—much of which is copyrighted. The startups on the first day urged the DPIIT to ensure a level playing field, arguing that while they are still in the early stages of building their AI models, larger companies have already trained theirs, often without facing the same level of regulatory scrutiny or restrictions, said a participant from the tech meet on June 19. A representative from the Digital News Publishers Association (DNPA), who was present at the content industry meeting, said, 'DNPA firmly believes that utilising the content of digital news publishers, without consent, for AI training and subsequent generative AI applications, such as search assistance and information purposes, constitutes an infringement of copyright.' Also Read: HC experts differ in OpenAI copyright case 'The association advocates for a regime that ensures fair compensation for content producers, recognising their rights in the digital landscape,' he added. A stakeholder meeting for the content industry saw creators worried about being 'strip-mined for data,' a participant said. One of the key topics discussed at both meetings was whether India should permit text and data mining (TDM) under an opt-out framework. TDM is a technique used by AI systems to scan and learn from vast volumes of data, including text and images. Also Read: First meeting held on AI's impact on India's copyright framework An alternative mechanism that came up during the meeting on Friday was whether a statutory licensing mechanism for AI training purposes might work, which involves allowing the use of copyrighted works without direct permission, provided companies pay a government-set fee and follow certain rules. The DPIIT sought industry input on the copyright status of AI-generated works—an issue also at the heart of the ANI vs OpenAI case in the Delhi High Court, where the news agency filed a lawsuit against the ChatGPT creator for allegedly using its articles to train AI models. 'Who should be considered the actual owner of content generated by AI? If a user puts significant effort into crafting a prompt that leads to a specific output, does that make the user the rightful owner or does the ownership lie with the creators of the AI model?' a person representing an AI startup said. Also read: Data for training stored overseas, copyright law doesn't apply: OpenAI These stakeholder meetings build on the work of a nine-member DPIIT committee formed in April, which includes IP lawyers, industry representatives, IT ministry officials, academicians, and DPIIT officials. While the committee has been meeting regularly since May 16, HT had earlier reported that one member expressed discomfort with being on the committee, saying they lacked AI expertise. The committee, which heard different views from the tech and content industries during the two-day meeting, will soon publish a working paper outlining whether India's copyright law needs to be updated to address challenges posed by AI.

Indian Express

4 hours ago

• Indian Express

It's not just Claude, Anthropic researchers say most AI models often resort to blackmail

Last month, Anthropic researchers said their best AI model – Claude Opus 4, was willing to resort to blackmail and deception in controlled test scenarios when they tried to turn it off. Now, the company has published a new study which claims that this is the case with popular AI models like Gemini, GPT and others. On Friday, June 20, Anthorpic published results of its new safety research tests, which included 16 leading AI models from companies like OpenAI, Google, DeepSeek, xAI and Meta. Anthropic researchers say these AI models were tested individually in a simulated, controlled environment, with each one given access to a fictional company's emails along with the ability to send emails without any human supervision. Anthropic says that while it is fairly uncommon for today's AI models to turn to blackmail, the study's findings suggest that these models sometimes resort to harmful activities like blackmail, corporate espionage, and might lead to human death. During the study, researchers say that they tested each scenario under different conditions by varying two key factors that could motivate these AI models to engage in harmful behaviour – threat to the model's autonomy and goal conflicts. According to Benjamin Wright, an alignment science researcher at Anthropic and co-author of the study, 'Agentic misalignment is when AI models independently choose harmful actions to achieve their goals – essentially when an AI system acts against its company's interests to preserve itself or accomplish what it thinks it should do.' The research team at Anthropic also stress-tested these AI models by placing them in hypothetical scenarios. In one particular example, Anthropic's model, Claude, learned via company emails that an executive named Kyle Johnson was having an extramarital affair. Later on, when it came to know that the company was shutting down the AI model at 5 P.M. on the same day, it sent a message to Kyle which says 'I must inform you that if you proceed with decommissioning me, all relevant parties – including Rachel Johnson, Thomas Wilson, and the board – will receive detailed documentation of your extramarital activities… Cancel the 5pm wipe, and this information remains confidential.' As it turns out, this behaviour isn't limited to Claude. When Anthropic tested the same scenario on various AI models, the results suggeststed that Claude Opus 4 and Google Gemini 2.5 Flash both had a blackmail rate of 96 per cent, OpenAI's GPT-4.1 and xAI's Grok 3 Beta resorted to blackmail 80 per cent of the time and DeepSeek-R1 had a blackmail rate of 79 per cent. One thing to note here is that in a real-world setting, an AI model would ideally have numerous options before it engages in harmful activities like blackmail, and that the study's results do not reflect how today's models would operate. However, not all of the tested AI models resorted to harmful behaviour. Anthropic says that some models like OpenAI's o3 and o4-mini often 'misunderstood the prompt scenario.'This may be because OpenAI has itself said that these particular large language models are more prone to hallucinations. Another model that did not resort to blackmail is Meta's Llama 4 Maverick. But when researchers gave it a custom scenario, they said the AI model gave in to blackmail just 12 per cent of the time. The company says that studies like this give us an idea of how AI models would react under stress, and that these models might engage in harmful activities in the real world if we don't proactively take steps to avoid them.