GPT-5 and Academic Research: Automating Writing, Synthesizing, Peer Reviewing and Data Analysis with AI

OpenAI’s GPT-5 was just released yesterday, and CEO Sam Altman compares the model’s development to educational milestones: GPT-3 was like a high school student, GPT-4 a college student, and GPT-5 is now a PhD graduate. What does this mean for academics? In this on-depth article, I’m testing how good GPT is at academic writing, synthesis, data analysis, and peer review. We will also dive into technical details and the way GPT-5 works.

GPT-5 Metrics and Features

The main difference you will notice when using GPT-5 is that you don’t have to choose the model anymore because GPT-5 contains multiple models and automatically chooses which one to use to answer your query. This is called a “real-time router”. Simpler queries go to fast models, complex queries go to a more powerful “thinking mode”. This can be a blessing and a curse, as most tasks are likely completed faster without overusing ChatGPT. However, you might not always get the desired output.

As Ethan Mollick points out, “it just does things.” And it’s sometimes not clear which model works in the background, leading to sometimes better, sometimes worse results. However, you can influence that by using the right prompt (more on this later).

GPT-5 recently achieved some very impressive results, like winning the gold medal in a (pre-college) math olympiad and coming in second in a programming contest. In the former, the AI also came up with a novel mathematical proof. Given that just a year ago, AI wasn’t anywhere near winning these contests, it is generally assumed that AI will start outperforming humans in the next year. (The models used in both cases were specialised ones and not the generic models you have access to.)

How much is GPT-5 better than GPT-4o and o3 versions?

GPT-5 is objectively better than the predecessors GPT-4o and o3. It scores higher on key academic benchmarks, like PhD-level science questions, math questions, and coding tasks. (One small caveat: the benchmarks are also developed by OpenAI). GPT-5 is also more reliable. Its “thinking” mode produces 78% fewer incorrect claims than o3, which is a significant improvement in factual accuracy.

You might think that benchmarks are one thing, but is it practically relevant?

In the rest of the article, we are going to dive into a few use cases, and you can answer that question for yourself. I asked others in the Effortless Academic community what they think, and here’s what one of them said:

To join the Effortless Academic community, simply sign up for any EA course or webinar. The community consists of hundreds of academics. We discuss new developments around AI tools, literature reviews, digital note-taking, and the future of academia.

Why is GPT-5 better?

Its value comes strongly from “product” track improvements: higher reliability, better instruction-following, and lower error rates. These qualities are more critical for daily research than performance in a single competition or domain. Importantly, OpenAI claims to have reduced hallucinations: GPT-5’s “thinking mode,” produces 80% fewer factual errors than GPT-4o.

But there are also technical improvements:

Nitty Gritty Technical Improvements of GPT-5 (For the Enthusiast)

OpenAI released a so-called GPT-5 system card, which is a very lengthy 60-page document containing all the technical evaluations and details about the new release. Here are some things I found in it that I think will be most relevant for academics. Feel free to skip that section if you’re not into technical statistics.

The document distinguishes between a “thinking” and “default” model, which in practical terms, just means the model takes longer to answer your queries.

1. Major Factual Errors Reduced by up to 78% in GPT-5

For academic work, this means fewer major factual inaccuracies. Depending on the version, GPT-5 generates 44% to 78% fewer responses containing at least one major error compared to previous models.

2. Hallucinations in GPT-5 reduced by up to 65%.

GPT-5’s thinking model has a hallucination rate 65% smaller than OpenAI o3, and the faster default model has a hallucination rate 26% smaller than GPT-4o in ChatGPT (with browsing enabled).
For academics, this means that you can rely more confidently on the factual accuracy of information generated by GPT-5, especially for research and drafting, reducing the need for extensive fact-checking. However, we are speaking of a reduction, not an elimination of hallucinations – remain vigilant when it comes to using it.

3. GPT-5 Accuracy on Complex Health Questions Increased to 46.2%

For researchers in health and medicine, GPT-5 offers specific, verifiable improvements. On the HealthBench Hard benchmark, which tests performance on complex medical questions, accuracy improves from a previous 31.6% to 46.2%. This is supported by an 8-fold reduction in hallucinations on these difficult topics. On urgent health queries, errors are reduced by over 50 times compared to some prior models.

4. Software Bug-Fixing Success Rate of GPT-5 Reaches 74%

In computer science, the models are more effective at solving practical software issues. On the SWE-bench Verified benchmark, which assesses the ability to fix real-world software bugs, the new model is the top performer. It achieves a 74% success rate in solving these problems on its first attempt, an improvement over the 68% from the previous version.

My take: I’ve always found AI’s impact in coding to be much more practically relevant than in other fields. It feels like we are a year or two ahead in coding tasks compared to other areas.Probably, also due to the better integration into everyday processes. like with GitHub Co-Pilot.

Use Cases For Academics: A Practical Exploration

Technical statistics mean nothing if they don’t have a practical relevance for your work. Let’s test GPT-5 on a range of AI-enhanced workflows specifically for academics: synthesising and writing, criticising, and some data analysis.

Use Case #1: AI-Assisted Academic Writing and Synthesis

Especially young career researchers always struggle with the blank page. The easiest way to overcome this is to save a bullet point narrative of your research and use AI to convert that into writing. This was part of my “Writing with AI” workshop (you can get the recording). With GPT-5, we can upgrade that workflow a little bit. I decided to upload 10 papers to GPT-5 that I know are relevant to the topic of my next paper. The prompt is deliberately crude and asks to write three paragraphs. I am being deliberately very vague to simulate how someone might approach it who has no idea about the subject matter:

I need you to write an introduction to an academic paper citing all of these papers. 
Write three paragraphs. 
First, I broadly about the importance of understanding biodiversity rearrangements. 
Second, something about the drivers of biodiversity rearrangements and range shifts
Third, something about the problems. 
Paragraphs shouldn't be more than about 200 words. They should start with a topic sentence and end with an important conclusion. Logically, they should flow from broad to narrow. If using citations, use them in line, not at the end of the sentence. Where possible, make your writing concise and clear without unnecessarily long sentences. 
Additionally, give me a list of the papers and a short one-sentence summary of what their main finding was. 
Think hard about this problem solve it step-by-step. 

Here is the result:

The most important part about the result is that GPT-5 extracted facts about the paper and rearranged them into the right context. The writing is not perfect, but a good prompt can easily improve it (See this tutorial for getting AI to write in your academic voice). The writing assistant I teach people in my Effortless AI course has about 20 times the instructions that I use for this prompt.

There is a small nuance that might have slipped your attention. The last sentence: “Think hard about this problem, solve it step by step”. In my experiments, omitting that sentence resulted in a generation time of about 10 seconds, compared to the 2.5 minutes it took for the result above. While the writing was not significantly different, the factuality and content greatly benefited from the thinking time. This demonstrates how GPT-5 chooses the model for solving your query and explains the difference stated earlier about GPT-5 thinking and default models. Our prompts need to evolve as well.

In conclusion, however, I’m impressed by how well GPT-5 gets at writing and sincerely convinced that at some point, academic writing will be something we do more for leisure or aesthetic purposes rather than necessity, similar to calligraphy that has once been the basis for printing books and is now relegated to the far corners of creative pursuits.

Use Case #2: Criticising and peer-reviewing your writing

This is a somewhat controversial use case, but an unavoidable one. More and more scientists are using AI to write their papers, resulting in a higher volume of papers in less time. The general policy today, however, is that AI should not be used for peer review out of concerns for data privacy and most journals already struggle to find peer reviewers. This puts pressure on the academic system and will likely lead to some form of reformation either through the use of AI or by incentivising peer reviewers in some fashion.

I wanted to test how good GPT-5 can be for peer-reviewing my own manuscript. And the results impressed me the most out of all use cases. I uploaded my manuscript (without figures) with this prompt:

<Copy & Pasted Manuscript>
This is a version of my manuscript, and I want you to take the role of a critical peer reviewer. Specifically, look for logical flaws, Extensions you can think of And any mistakes you might notice. Point out if something is written more verbosely than it could be. 
Be critical but practical, minding the scope of the manuscript (For example, I could explore the rate of climate change in New Zealand, which would be very much related to this manuscript, but it would be completely out of scope.)
Make 20 concrete suggestions, writing them with a header each and a short one- to two-sentence summary of what you think should be improved or is the problem.
Be diligent and think hard about this problem.

The result is, in many ways, mind-blowing. GPT-5 completely understood what I wanted and revealed a lot of very small inconsistencies that I would have missed myself.

While the majority of these things will be too specific for you to appreciate. Let me guide you through a few ideas that I find completely mind-blowing:

• GPT-5 is catches logical inconsistencies (point 2). Each species I analyse in my work gets a score, and every species with a score below 0.7 gets discarded from the analysis. Yet at one point in the discussion, I mentioned that the average score was 0.62, which technically cannot be, since all scores must be above 0.7. That tiny detail requires considerable reasoning, especially since the 0.7 threshold is discussed in the Methods, while the 0.62 number appears down in the Discussion. (The real culprit was a typo, and the average score was supposed to be 0.82).
• GPT-5 adds external knowledge to the peer review (point 9): In my research, I’m generating so-called feature importance scores for some variable. These scores, however, suffer from a few drawbacks. The (better) alternative is called: Permutation importance scores, which GPT-5 refers to but is not mentioned in the manuscript.
• GPT-5 is detail-oriented (point 15): I accidentally linked to the wrong figure in my results, which AI has noticed and reported. Surprisingly, three people have read over this manuscript and have not seen that problem.
• GPT-5 is technical (point 19): Here, the AI refers to “spatial autocorrelation”, a topic that is not mentioned in the manuscript but often affects models like the one presented in my paper. Addressing it is likely a valuable suggestion and might come up later.

In summary, I think GPT-5 did an outstanding job at critiquing my work, finding small, detailed flaws that many humans overlook, and also made technical, large-scale suggestions and pointed out bigger-level issues. With a more detailed prompt on how to review the paper and sort the output, I think it would generate even more outstanding results.

The whole process took 54 seconds.

This being said, I don’t see how, in the very near future, a lot of peer reviews will not be fully AI-generated or at least pre-screened by AI by editors.

Use Case #3: Complex data analysis and statistical tests.

I have not met an ecology student who would not be struggling with some form of statistics or programming, or at least not be slowed down by these things. (Myself included!)

In this test, I decided to give GPT-5 a huge data file (~190 MB csv file), which takes Excel almost a minute to open and ask it to generate some statistical analysis and data visualisation on it. The file contains climatic data (19 variables) from about 5000 locations across New Zealand over a 70-year period. My question is, how did the climate change at this time? And some description of how I would like to visualise that. This is effectively what I did for one of my papers, which became one of the four results, so it is a practical exercise, not just a theoretical one. I made a video for the result:

The challenge here is that the data file is extremely long. Most operations, even on my machine, will take quite a while to complete. It should also be noted that GPT-5 is not explicitly made for this, but can do it as part of its general skill set. If you want a more dedicated data analysis AI, try Julius. The result is definitely what I wanted and is also a small proof for climate change. Here is what Dr. GPT found:

This plot just compares the average climate in the 50s and 60s to the 2000s and finds a statistically significant increase in, for example, temperature.

I think what this use case visualises is that science will become more and more an idea business because I don’t have to think hard about what types of resources I have to investigate, which questions I can investigate; all of them by delegating them to AI. Obviously, the results need to be vetted, but writing 300 lines of code that process a massive data file is not something I could have done even if given ten times the time.

Outlook And Summary: The Reshaping of the Academic Landscape

GPT-5 is a tool that will reshape academic research. I hope my use cases here made that palpable. Its capabilities in synthesis, analysis, and simulation can break down barriers and open new lines of inquiry. Here is how I think it will influence the future:

Academia as An Idea Domain

The more that (especially young career researchers) can eliminate the tedium of writing, the more ideas they can explore. I do not particularly enjoy academic writing and find it a waste of time. But I’m absolutely fascinated by analysing data and synthesising information to coming up with new concepts. So far, however, I have spent about an equal amount of time on writing and doing research. The writing part is becoming increasingly automated, while the coding part is becoming more efficient. In my opinion, this makes us scientists more and more idea and big-picture people, which is very positive for me personally.

Interdisciplinary Bridge-Building

Academic disciplines are often separated by their own languages, jargon and methods. GPT-5 can act as a bridge between them. By synthesising information from different domains, it can identify connections that a human specialist might miss, but not because they are incapable, just because of the way academia is built. A psychologist usually doesn’t meddle with the methods used by a machine learning specialist. This will hopefully unify fields.

The Qualitative Research Revolution

Qualitative research traditionally involves the manual analysis of unstructured data. GPT-5 can change this process. It can perform advanced thematic analysis on large volumes of text, video and audio, identifying themes and patterns at a scale that was previously not possible. Just think at how detailed my peer review example was to see the attention to detail that GPT-5 brings to the average academic.

The Conductor Mindset: Redefining the Role of the Academic

Expert-level AI does not make academics obsolete. It requires an evolution of the academic’s role. You shift from being a musician, who performs a task, to a conductor, who directs AI systems that perform tasks. Same medium, different skillset. I am talking about this in the Effortless AI course, that was just updated in 2025 (and will recieve another update with GPT-5 soon):

From Soloist to Conductor

In this new model, your primary value is in the strategic design and oversight of research. The key skills are asking the right questions, designing experiments, and critically evaluating the AI’s output. This is the “expert in the loop” model, where your expertise guides and validates AI-generated work to ensure quality and rigour.

This change requires a new set of skills. Rote memorisation becomes less important than critical thinking, creativity, and adaptability.

The most successful academics will be those who can collaborate effectively with AI, using it as a tool to amplify their intelligence. That’s possible with custom GPTs and AI agents. Learn more in my free AI lit review & writing course (link below!).

Frequently Asked Questions on GPT-5

Is GPT-5 coming out for free?

Yes, a version of GPT-5 is available for free to all ChatGPT users, but with usage limits. Free users can send a limited number of messages (around 10 every 5 hours) before the system switches to a less powerful model. For higher usage limits and access to the most powerful versions, you need a paid subscription like Plus or Pro. (The ChatGPT subscription is, in my opinion, the most cost-effective of all academic tools, as it can be modified to have the behaviour of specialised tools, avoiding their additional subscription fees.)

Where can I get GPT-5?

You can access GPT-5 through the official ChatGPT website (chat.openai.com) or the official mobile apps. GPT-5 is the new default model for all users, so no special action is needed to enable it. However, the rollout is happening step-by-step. If you’re reading this article immediately after its release, you might not be able to see it yet. In that case, wait for a few days, and it should just be enabled.

How much better is GPT-5 than GPT-4?

GPT-5 is a significant improvement over GPT-4. It is faster, more accurate, and hallucinates less. It uses a new “unified” system that automatically switches between a fast model and a deep reasoning model. Sam Altman compared the improvement to Apple’s introduction of the Retina Display, calling it a substantial upgrade. You can find nitty-gritty technical details in this lengthy PDF that OpenAI published along the GPT-5 release.

GPT-5 vs Claude 4

GPT-5 and Anthropic’s Claude Opus 4.1 are two leading AI models and are close competitors, especially in tasks like coding.

• GPT-5’s Advantages: Its main strengths are its versatility and lower cost. It has strong multimodal capabilities (text, images, voice) and a broad feature set in the ChatGPT interface. Its API is significantly cheaper than Claude’s, making it more accessible for developers.
• Claude 4.1’s Advantages: Claude is often noted for its precision in complex, multi-file coding tasks. It is also regarded for its safety features and transparent “thinking” processes, which is useful for enterprise workflows.

For most academic users, GPT-5’s feature set and unified system make it a versatile assistant. For specialised, enterprise-level coding, Claude 4.1 is a strong alternative. However, if you already have one subscription, just stick to it. The differences are often minimal.

GPT-5 vs Google Gemini 2.5 Pro

Both models claim to have achieved human-level performance in mathematical benchmarks and will likely be sufficient for the average user. However, a significant advantage that Gemini has over GPT-5 is that it is connected to your Google Drive and can access your PDFs and files there. If you store your papers inside Google Drive, Gemini can act as a personal search engine and knowledge synthesiser for the documents that you are reading. At least for academics, this is a very attractive use case that is much harder to achieve in GPT-5. In my experience, however, GPT-5 does a better job at academic writing than Gemini.

Is your PhD obsolete with GPT-5?

This is a question many people might ask themselves. I think the answer is “not at all”. However, the mindset we have needs to shift from being someone who does research (including academic writing, reviewing, and coding) to someone who investigates ideas and does not care so much about how they are written down or the methods used to investigate them, trusting that another entity (in this case, AI) can do it better and faster. This will entail significant changes in the academic landscape, especially as AI increasingly accelerates publishing, and fewer and fewer peer reviewers are available to read these papers. We will likely see automation in areas that we would have never thought possible, which will transform academia ultimately into a pure idea domain.