OpenAI introduced GPT-5 with bold language and sweeping promises. In their launch announcement, the model was described as a unified system capable of delivering “expert-level responses” across domains from coding to writing to healthcare—setting the expectation of a PhD in your pocket. The media amplified this framing. Sam Altman told reporters that using GPT-5 “feels like talking to an expert with a Ph.D., no matter what topic you bring up.” Reuters echoed him: GPT-5 was pitched as something you could ask “a legitimate expert, a PhD-level expert, anything.”

It’s evocative language, but also a very high bar. And I’ve grown wary of big labels. “Revolutionary.” “Autonomous.” And now, “PhD-level expert.” It sounds good in a press release; it breaks in a lab meeting. So last week we set out to test the claim, not to crown or cancel GPT-5, but to calibrate it.

We stress-tested it on a real genomics problem and asked practitioners who live in code and analysis every day what it actually feels like in their hands. GPT-5 framed the problem cleanly, proposed plausible experiments, and even anticipated failure modes. That felt like a strong doctoral candidate in the room.

Then it overstated a few claims and cited with more confidence than care.

That reminded me: expertise isn’t just knowing what to say, it’s knowing how hard to say it.

This piece calibrates the language and the promise. GPT-5 can feel “PhD-level” when the work demands design, critique, and prioritization under constraints. It drifts when we reward it for sounding right instead of being right.

The Operating Thesis: GPT-5 Mirrors the Operator

Here’s the opinion I’ll stand behind: GPT-5 is an intelligence mirror. In expert hands, it looks PhD-level because it reflects a strong mental model; in vague hands, it reflects the vagueness.

Rajdeep Mondal, Senior Data Scientist at Elucidata, put it plainly:

“It’s a force multiplier. In the hands of a senior engineer or analyst, it’s 10–15×. In the hands of a novice, it’s often underwhelming.” - Rajdeep Mondal

That line matters more than any benchmark. Capability scales with the person at the keyboard. If you bring domain context, standards, and a sense of trade-offs, GPT-5 compresses weeks of whiteboarding into hours. If you don’t, it’s still useful, but rarely decisive.

Rajdeep also described a healthy decision habit:

“I triangulate across models, then decide. It’s a friend who knows a lot, but I’m still the one who ships.” - Rajdeep Mondal

In other words: use it to widen the option set, not to outsource judgment.

What “PhD-Level” Should Mean (and Where GPT-5 Fits)

A PhD-level contribution is not encyclopedic recall. It is the ability to:

• translate a messy, ill-posed question into tractable sub-problems,
• propose discriminating experiments with clear decision criteria,
• anticipate failure modes and salvage paths,
• qualify claims so verbs match evidence.

On these, GPT-5 performs well when you press it correctly. In our trials, it built decision trees, sequenced work sensibly, and identified controls. Where it stumbled was source rigor and degree of certainty, the places committees live.

Prompting Styles: What Actually Worked

Arushi Batra, Ph.D. Marketing Associate at Elucidata, compared two styles across real tasks (gene-list pulls, workflow design, experiment planning):

Approach 1 – Conversational, human-like prompts

Very natural, “messy” style - like talking to a colleague mid-thought.

Pros: Mimics real-life collaboration; tests AI in realistic research scenarios.
Cons: Less structured, so AI sometimes misreads context or misses precision details.

Approach 2 – Detailed, context-rich prompts

Highly structured and data/context-heavy.

Pros: Provides AI with clear boundaries; better for generating workflows, gathering data, or offering alternative perspectives.
Cons: Can still hallucinate; human review is necessary for feasibility and regulatory constraints.

What we saw: Structured prompting consistently produced more actionable output. Conversational prompting was great to open the space, weaker at closing the plan.

“GPT-5 is fast, versatile, and insightful, but not a full-fledged co-scientist.”- Arushi Batra

This is how GPT-5 output quality differs on the basis of the quality of prompts-

Where human intervention is needed in both these approaches?

• Approach 1: Check references/dataset context; fix database trust issues (e.g., ClinVar); translate concepts into implementable steps.
• Approach 2: Validate feasibility (regulatory, sample throughput); correct hallucinated conclusions; turn “fresh lens” into executable workflows.

Arushi’s takeaways-

• Human oversight is non-negotiable; hallucinations and context slips persist.
• Structured, context-rich prompts outperform conversational ones for design and delivery.
• Today, GPT-5 is best framed as a brainstorming assistant and data aggregator, not an independent PhD-level collaborator.
• Best practice: Explore with conversational prompts → formalize with structured prompts → always validate assumptions and constraints.

“Yes, it pulled a gene list quickly and suggested obvious objectives. But when it came to designing experiments, it stalled, unless I explicitly nudged it toward gaps and alternatives.”- Arushi Batra

The View from the Trenches

• Mirror, not substitute. GPT-5 reflects the user’s clarity. The ceiling is your problem framing.
• Structure is disguised expertise. Conversational prompts open the space; structure closes it. If you can structure the ask, you’ll get “PhD-like” output.
• It reasons with you, not for you. Ask it to pit mutually exclusive models against each other and to pre-declare decision criteria; left alone, it will still overstate.
• Keep human ownership of claims. Confident prose ≠ verified citation. Treat references as leads, not verdicts.

Bottom Line

If “PhD-level expert” means anticipate gaps, challenge assumptions, and spot non-obvious next steps without being spoon-fed, GPT-5 isn’t there yet.

Think of GPT-5 as a sharp, hardworking new grad: fast and thorough, but not an independent thinker. With a strong lead, it makes good teams great; on its own, it still needs direction and review.

For further perspectives on how AI is being applied in life sciences, explore our webinars, whitepapers and case studies on real-world use cases.

By

Nidhi Khurana, Ph.D.‍

Content Marketing Manager

Elucidata