OpenAI forum reveals how deep research transforms inquiry
OpenAI has introduced a new agentic AI system called 'deep research,' designed to handle complex, time-consuming research tasks by simulating the work of a human analyst. Presented by researchers Isa Fulford and Edward Sun during an OpenAI forum event, the new tool is powered by a fine-tuned version of OpenAI's upcoming O3 model and leverages advanced reasoning and browsing capabilities.
"Deep research is an agent in ChatGPT that can do work for you independently," Fulford explained.
"You give it a prompt, and it will find, analyse, and synthesise hundreds of online sources to create a comprehensive report at the level of a research analyst."
The system is intended to help users across a range of sectors—from academia and medicine to business and software development. "Members are finding that deep research asks clarifying questions to refine research before it even starts," said Fulford. "We think that deep research can accomplish in tens of minutes what would take a human many hours."
The model represents a major step forward in OpenAI's work with reasoning systems, building on reinforcement learning techniques introduced in its earlier models. Fulford explained how the company developed the tool: "We launched O1 in September of last year. This was the first model that we released in this new paradigm of training where models are trained to think before answering… and we called this text where the model is thinking, 'chain of thought'."
This method of structured, internal reasoning proved effective not only in tasks such as maths and coding, but also in navigating complex real-world information environments. "Around a year ago internally, we were seeing really great success… and we wondered if we could apply these same methods but for tasks that are more similar to what a large number of users do in their daily lives and jobs," Fulford said.
Sun detailed how the tool works by combining reasoning with specialised capabilities like web browsing and code execution. "The browser tool helps the model to aggregate or synthesise real-time data, and the Python tool is helping the model to process this data," he explained. The system dynamically alternates between reasoning and action, using reinforcement learning to improve over time.
One striking example involved analysing medal data from the 2020 Tokyo Olympics. "You can see how the model interleaved reasoning with actual tool calls to search for information, refine the data, and process it programmatically," Sun said.
Unlike older approaches that rely on a single-pass search or instruction-following, deep research iteratively refines its answers. "We train the model with end-to-end reinforcement learning," Sun added. "We directly optimise the model to actively learn from the feedback, both positive and negative."
OpenAI tested the model extensively against both public and internal benchmarks. According to Fulford, "the model pairing deep research scored a new high of 26.6%" on the Humanities Last Exam, an expert-level evaluation spanning over 100 subjects.
On another benchmark, GAIA, the tool also achieved a state-of-the-art result for multi-step web browsing and reasoning.
The model also underwent safety evaluations prior to release. "We did extensive red teaming with external testers, and then also went through preparedness and governance reviews that we always do at OpenAI," Fulford said.
Despite strong results, the researchers acknowledged current limitations. "It still may hallucinate facts or infer things incorrectly," Fulford said.
"Sometimes it struggles to distinguish between authoritative sources and rumours."
Use cases continue to emerge in unexpected domains. "People might be using the model a lot for coding. And that's been a really big use case," Fulford observed. Other domains include scientific and medical research, where professionals have begun verifying the model's output against their own expertise.
Users are also adapting their behaviour to suit the model. "We've seen interesting user behaviour where people put a lot of effort into refining their prompts using O1 or another model," Fulford said. "And then only after really refining that instruction, they'll send it to deep research… which makes sense if you're going to wait a long time for an output."
Currently, deep research is available to users on the Plus, Pro, Teams, Enterprise and EDU plans.
"We're very excited to release a smaller, cheaper model to the free tier," Fulford confirmed. The team also plans to improve personalisation and explore ways to let users incorporate subscription services or private data into the research process.
"This showcases how the model can effectively break down a complex task, gather information from various sources, and structure the response coherently for the user," Sun said in closing.
OpenAI's forum audience, composed of members across academia, government, and business, left the event with a clear sense that deep research marks a meaningful step toward AI systems capable of handling work currently done by skilled analysts.
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
The Spinoff
9 hours ago
- The Spinoff
‘I wanted to be a teacher not a cop': the reality of teaching in the world of AI
As AI becomes more ubiquitous, teachers across the board are struggling to adapt. Hera Lindsay Bird reports. *All names in this article have been changed. Julia*, a New Zealand teacher working in an American university, doesn't explicitly forbid her students from using AI. The top ranked public university where she teaches academic writing has told her not to. Instead, the students in her classroom must abide by two simple rules. They must ask for permission. And they must provide the LLM prompts, transcripts, and a written reflection on how AI contributed to their work. Of her class of approximately 20 students, several have acknowledged using ChatGPT. It's possible others have managed to sneak AI past her undetected, but she's not so worried about that, because it means 'on some level they got the point of the class, which was about adapting writing to different audiences, and as the audience for their assignments I was an openly seething AI hater.' But there are two repeat offenders who continue to hand in essay after essay of LLM generated writing. The essays are 'bloviating nonsense,' complete with fake quotes and made up sources. The two students repeatedly dodge her requests for in-person meetings. They only reluctantly agree to meet through Zoom. When she questions them about their use of ChatGPT, they lie, saying they only ever used it as a brainstorming tool. Only when she manages to 'circle them in conversation' do they grudgingly give her a fuller account of how they produced the work. Julia is sympathetic to the pressures her students are under. She's patient with them. She painstakingly walks them through the assignment step by step, offering suggestions for how they might resubmit their work. 'What was really sad was the fact it didn't seem to stick,' she says. 'I got final AI assignments in the end, and had to again chase them for versions they had at least edited using their own language, righteously incandescent with frustration like Hades in Disney's Hercules.' She passes the students with low marks. She's not convinced the university would back her up if her students contested their failing grade. 'My main beef with AI is that it made me into a grown adult asshole who had 18-year- old enemies,' she says. 'I wanted to be a teacher not a cop.' In the last few years LLMs and generative AI has graduated from a novelty tool that generates attractive women with more than the amount of medically recommended thumbs, to something that can write a convincingly mediocre essay on the role of ambiguity in Henry James's The Turn of the Screw. Two weeks ago, James D. Walsh published a piece in New York Magazine called 'Everyone is Cheating Their Way Through College' about how students and teachers in the US college system are dealing with education in a post LLM world. The article painted a grim picture about the ubiquity of students using generative AI to cheat on their assignments, claiming, as per the subtitle, 'ChatGPT has unraveled the entire academic project'. New Zealand universities are facing similar problems. In May, Victoria University of Wellington made the decision to enforce a handwritten exam policy for two third year law papers in an attempt to prevent student cheating. Days after the announcement, the university backpedalled. The Post reported that, in response to student concerns, the university had settled on a compromise. Half of the exam would comprise a handwritten multi-choice test. Students would still be permitted to use their laptops for the second half of the exam. Use of AI was still strictly forbidden and would be 'enforced through active and enhanced invigilation.' Universities aren't the only ones struggling with the new reality. Almost every educator I spoke with, from primary school teachers to those supervising postgraduate dissertations, raised serious concerns, with some teachers estimating that up to 80% of their students relied on ChatGPT to complete assignments. I spoke to MA supervisors whose history students theses were riddled with fictitious sources and 'archival' Midjourney photographs, and primary and intermediate school teachers, who said students as young as 11 were using it to answer simple personal prompts, such as 'what did you do in the summer holidays?' and 'what was your favourite penguin in the text?' All of this raises the question: What's the point of an education? Is this the inevitable result of an increasingly consumerist tertiary sector, where students are simply paying for credentials in exchange for entry into the job market? Should we learn to embrace AI as an educational inevitability, in the way that we've come to accept the pocket calculator? Are we being unnecessarily catastrophist? Ben*, an English teacher at a North Island all-boys school, says 'I find it really frustrating to read boosterish pieces about it in the press because they all assume that students will use it to learn. AI is seen as offering kids autonomy over their own learning, but this is not an autonomy they know what to do with! Students are using it to solve their problems and get answers, not to think with.' Ben worries that information literacy is falling by the wayside. 'Traditional research and evaluation skills are totally bypassed by AI,' he says. 'If [a student] asks 'what is evidence about how veganism is unhealthy', it will supply them with exactly that, without them having to do any sleuthing, comparison or evaluation'. The greatest concern from teachers was the atrophying effect it was having on students' critical thinking capacities. Julia from the undergraduate writing programme says, 'They can't get their head around why authorship matters, so they can't tell slop from scientific journals.' This problem is further complicated by the fact that 'reputable journals are accepting GPT bollocks that slips through'. Lily*, a lecturer in the arts at a large New Zealand university, predicts that her colleagues will be shocked come exam season. 'There is going to be a vast difference between what has been submitted for take home assignments and what students will be able to produce in exams.' There is an argument to be made that ChatGPT can level the playing field, especially for students who struggle to translate their learning into academic language. But as other teachers were quick to point out, over-reliance on this technology actually has the potential to widen the gap. Mark*, a special education teacher working at a US high school, uses the example of a student with a reading disability. 'If a student has difficulty reading, difficult reading frustrates them. They can just tell the bot to read the article and answer the questions. They'll get the correct answers… and of course their special ed documentation will say they're making progress when actually they aren't'. Ben is adamant that AI is a tool being utilised in unhelpful ways. 'When all students have AI, why would they choose to do the hard task when they can make it three times as easy in an instant? Again we assume students understand learning in a more holistic sense and that they know what's best for their own academic progress. But that just isn't true. Students will (like all humans!) take the easy route out most of the time, and by enabling this, they're missing out.' So what, precisely, are teachers supposed to do about this problem? Unlike the plagiarism detectors of the previous academic generation, there doesn't seem to be any surefire way to prove that LLMs have been used, even though most teachers felt confident in their ability to detect the linguistic hallmarks of ChatGPT, citing essays that sound like 'management reports' chequered with phrases like 'these themes converge' or 'a conceptual alignment is revealed'. One teacher noted the 'degrassi era syntax'. Preventing students from using AI is becoming increasingly difficult, as the technology becomes more ubiquitous. Even if teachers ask for handwritten essays, or check over a Google document's revision history in an attempt to rumble obvious copy paste jobs, students are quick to adapt. 'They could get AI to rewrite it in the style of a 15-year-old, complete with spelling errors,' says Ben. Or painstakingly copy the LLM's essay by hand. So far, the only reliable method of preventing students from using AI is to revert to technology-free classrooms, with handwritten or oral examinations, but even this solution is potentially short-lived, as technology like Google Glass becomes more mainstream. Some teachers have gone to great lengths to identify students using AI, like the lecturers in the New York Magazine piece, who hid 'gotcha' instructions in their assignments, such as including 'mention Ariana Grande' in white font to identify students who pasted the assignment brief directly into ChatGPT. But strategies like this have a limited shelf life, as students communicate and learn to adapt. The efforts of teachers to prevent the use of LLMs are often undermined by the educational institutions themselves. Many of the assessors I spoke to said that even when they had identified something they were 99% sure was written by generative AI, they were often powerless to do anything about it. Some teachers were explicitly discouraged from banning AI. Others had banned AI, but the policy was effectively unenforceable, as the use of AI is notoriously hard to prove, and the universities feared complaints or litigation. Many teachers I spoke to felt as if they had been left to navigate the grey areas alone. Unsurprisingly, many teachers are becoming increasingly disillusioned. Julia said dealing with AI took up half of her mental load, 'due to the cat-and-mouse of it all'. Another teacher, Jo* who has been teaching undergrad at a large New Zealand university for around a decade, says policing it is a nightmare. 'It takes up so much time and our workload calculations haven't changed to account for this'. It's not just a physical toll for Jo either. 'Dealing with student work which is entirely AI written is by a wide margin the most discouraging thing I've experienced working in academia, and given how many very discouraging things have happened lately that's really saying something.' Concerns over AI aren't limited to educators. I spoke to a few students, who were dismayed to discover their essays were being marked by AI. 'There is a trend for teachers to use it for marking but I'm not a fan,' says Ben. 'Marking is about finding out where your students are at so you can sequence your lessons. Taking the teacher out of the equation is antithetical to good teaching. It's like driving with your eyes closed.' Other teachers have begun using AI as a lesson planning tool. For some, this is an effective time-saving measure and eases the burden on a sector that is notoriously under-resourced and overworked. As Ben is quick to point out, teachers using AI to plan lessons is hardly equivalent to students using it to answer questions. Teachers, he points out, have the competence and skill base required to use AI 'selectively and thoughtfully'. But the outcomes aren't necessarily positive. One witness reported a new entrant teacher getting students to write descriptions of monsters, before entering the descriptions into ChatGPT to create illustrations. But the generated pictures were inappropriately terrifying. To quote the observer; 'Why cant they draw the fucking monsters?' The news isn't all bad. I spoke to some teachers who said they'd observed positive use of LLMs by students to further their learning, like a Romanian sociology lecturer who saw students practising statistical analysis by generating fictitious interview subjects. Others, like Rachel*, a teacher at an international primary school in Malaysia, say that LLMs are a useful tool for translation. 'For some kids, it has really improved their creativity. If it's used effectively and taught well, it could revolutionise teaching.' But by and large, the mood was depressing. Are we staring down the barrel of the future, in which we're simply teaching one robot how to talk to another? Ben, the high school teacher, suggested a 'de-teching' of the classroom was inevitable, and that he and his colleagues have already seen positive interpersonal results from the government's school cellphone ban. 'We need to start thinking about what skills we want humans to have, regardless of their utility in an economy that will slowly be replaced by AI. There's no point trying to predict the jobs of the future. Let's think: what are the things we want everyone in our society to know and understand? What are the key skills, human and disciplinary, that we want everyone in our society to work on? What kind of traits and characteristics do we want to encourage for the good of the individual and the community? That's where we need to start,' says Ben. 'It's in our hands to make a more robust and more human education system. In fact, we have to. If we don't, and if we allow kids to outsource to AI how to think, write, imagine, and work, we'd be doing them a generational disservice.' Julia, despite her struggles, is quick to caution against hyperbole. She believes claims that everyone is using AI are 'highly exaggerated', and the people pushing these products are incentivised to promote the idea that 'not to cheat is to get left behind'. Most of her students, she says, are 'honestly trying to get by on their own thought and expression. So it's certainly not a lost battle we should just give up on due to perceived ubiquity of the tech'. Most teachers agree that the Rubicon has been crossed, and complete abstinence from AI tech is no longer possible. But providing students with the tools to think critically and parse fact from fiction is perhaps more urgent than ever, in a world that is increasingly full of bad actors and disinformation. As Julia says, 'eventually they'll need to rub two brain cells together in a circumstance where the stakes feel real to them, and the less they've practiced thinking and expressing their thoughts the more cringe their failures will be.'
NZ Herald
9 hours ago
- NZ Herald
Small Business On the Up: Former Olympian Willy Benson's PortaSkip journey in Hawke's Bay
I approached a business mentor who turned into my business partner, just pitching ideas and growing my knowledge. I was running a swimming club at the time and wanted something on the side that was a long way away from swimming. And then I did a few feasibility studies, went back and forth and was pretty happy to come up with PortaSkip. I saw something similar in Australia. The guys at Bunnings hire out skips, but they were bins on trailers that had to be forked over the back of the truck. Wanting to run a small business, we didn't want to have to invest in that sort of capital, so we came up with our skip bins that have a hydraulic ram underneath. It just requires a ute, and it can run pretty simply, meaning we could fill a gap in the market that wasn't really there in terms of something a little bit more nimble and mobile. PortaSkips are perfect for homeowners and landlords who are clearing out a property that requires getting into a tight space. How was the transition from a life as an Olympian to running a business? I was lucky that while I was training for the Olympics, I was also at Massey studying a business degree and ended up majoring in business management. So I'm probably one of the few people who actually got out of university and managed to use their degree. I didn't really know what I wanted to do when I stopped swimming, but I was lucky that I had the tools to go and find things that interested me and challenged me. I think especially when you're starting a new business, it's a little bit like my swimming, I had a goal to represent New Zealand, and that was four years away, but you've just got to keep chipping away for something. It can seem like a bloody long way away in the future, but if you trust the effort and you're genuine and you keep chipping away, there's a chance that you'll get to where you want to get to. I think there's a pretty strong crossover between starting a new business and your sporting goals and dreams in that regard. What was it like working during Cyclone Gabrielle? It's still unbelievable every day when you drive down the expressway thinking about that thing under water. That was tricky, certainly. My family home that I grew up in was under water, so we're digging silt out of mum's place and pulling out a bunch of our possessions that we grew up with. Meanwhile, we were running around like maniacs trying to turn around as many skips as we could for everyone trying to get stuff out of their homes and businesses that were under water as well. It was wild, wild stuff and hopefully stuff that I'll never have to see again. I think honestly coming out the other side of the cyclone and just sitting back and looking at what we were able to do for people, the amount of stuff we moved and being able to see people at the lowest point in their lives and know that you played a little part in just making that super tough time a little bit more manageable for them, I think that's been a big highlight though. PortaSkip co-founder Willy Benson remembers trying to supply bins for everyone in need following Cyclone Gabrielle in Hawkes Bay. Where do you want the business to be in the next few years? I think the biggest piece is just expanding it to other areas, like Palmerston North and Wellington, and to keep it growing. With any business, it's a challenge to get established, and then once you have an established business, it's like tough to keep that momentum and keep it going and not fall off or just plateau. I think the goal over the next two to five years is to have the business throughout New Zealand. What would be your advice to other budding entrepreneurs wanting to start a business? Just do your homework initially and make sure that everything lines up, and then just work hard and trust yourself. There are times when growth is pretty slow going, but if it's genuine and you trust yourself and you've got a good plan behind you, you're most likely going to come out the other side where you want to be. Having a mentor and having someone to bounce ideas off of and help vet things is massive too, whether that's experience from previous jobs that you can translate into a new business or just along the way having another person to help you troubleshoot when you've hit a wall, I think that's massive as well. Tom Raynel is a multimedia business journalist for the Herald, covering small business, retail and tourism.
Otago Daily Times
a day ago
- Otago Daily Times
AI use intensifying scams: Netsafe
Artificial intelligence (AI) is enabling fraudsters to devise ever-slicker romance scams, Netsafe says. The online safety agency recently presented updated resources as part of its Get Set Up for Safety programme, aimed at protecting older people from an upswing in sophisticated digital cons. Business development manager Sarah Bramhall said scammers might spend weeks or months building online relationships before seeking money. "Scammers most often use the techniques or the emotions of trust, fear and hope, usually in a combination. "So they will tap into human emotions." Exploiting lonely or companionship-seeking victims, scammers try to stop them sharing information with friends or family. "They will try to keep them isolated so that they don't tell anyone, because obviously otherwise friends and family will pick up on something happening." At some point the scammer will begin requesting money, sometimes large amounts or gradually increasing amounts. These requests could be couched in ways that played on people's natural desire to be kind or helpful. "Usually it presents itself in something like a medical requirement, they need to travel, they have got family that are sick. "Those sorts of things that really play on emotions." Kind-hearted people who felt they had developed a bond would feel like they wanted to help that person out. "Most of the time, people really don't recognise that they are being scammed in those scenarios. "It is really quite hard for even support workers and family to get them to come to that realisation because they suffer heartbreak, essentially." Generative AI tools were enabling scammers to polish their English, generate fake images or create believable back-stories. Poor grammar or language used to be a red flag that it was a scam message. "That is getting harder to pick up on now," she said. While there were many ways AI was opening up useful and beneficial possibilities, it was important to be mindful of some of the drawbacks of AI, in particular large language models such as ChatGPT, which could create "hallucinations" that could seem plausible but were falsehoods. "I just say 'sometimes AI can lie'." Netsafe has refreshed its portfolio of resources that can help organisations and individuals navigate the online digital realm safely. The material tackles challenges such as spotting scams, safer online dating, privacy settings, securing accounts and verifying requests for personal information. Get Set Up for Safety offers a wide range of resources, including checklists, fact sheets, videos and interactive activities. • To find out more, visit
