What prompt gets a good answer from a science AI?
Paste this into Claude, ChatGPT, Gemini or Whizi and fill the brackets. It sets the level, the sources and the format, which a bare science question leaves out.

Answer this science question: [your question]. I am [who you are, such as a ninth grade student or a first year biology undergraduate]. Search the web and use only peer reviewed papers or university and government pages. Explain it in [number] sentences, then list each claim with the paper or page behind it, its authors and its year. Mark any claim you could not source as unsourced, and tell me what you are least sure of.
Each bracket closes a gap. The level stops a graduate answer landing on a child. The source rule keeps a blog out of the evidence. The last line makes the model show you where to look first.
What does a good answer look like? Three worked examples
A good science answer states the mechanism, shows the working with units and tells you what it could not check. These three are written for this page. They describe what to look for and are not the output of any named model.
Example 1, a concept. Prompt: "Why does ice float on water? Explain it to a ten year old, then give the one sentence a physics teacher would add."
A good answer says that water molecules form an open, hexagonal pattern as they freeze, which spreads them further apart than in liquid water. Ice is therefore less dense than the water around it. The teacher's line names hydrogen bonds. If the answer says ice is lighter because it is cold, it is wrong.
Example 2, a calculation. Prompt: "How much energy does it take to heat 2 kg of water from 20 degrees C to 80 degrees C? Use a specific heat of 4.18 J per gram per degree C. Show units on every line."
A good answer converts 2 kg to 2,000 g, multiplies 2,000 by 4.18 by 60 and reports 501,600 J, or about 502 kJ. The check is that the unit on the last line is joules. If the units never get converted, the answer is a thousand times out.
Example 3, a paper search. Prompt: "Find three peer reviewed papers on whether [intervention] changes [outcome] in [group]. For each, give the authors, journal, year and DOI, and say whether it is a trial, a cohort study or a review."
A good answer gives three papers you can look up, labels the study type, and says outright if it found fewer than three. A confident list with no DOIs is the red flag. Open each link before you cite anything.
Is there a free AI for science questions and papers?
Yes. Consensus, scienceOS, Elicit and Semantic Scholar all have a free tier, and ChatGPT, Claude and Gemini do too. The table lists what each says on its own page as of November 2026, with seven tools plus Whizi. Where a page was silent, the cell says not stated.
| Tool | Free use | Sign up | Paid price | Sources and papers | Best at |
|---|---|---|---|---|---|
| scienceOS | Free plan at 0.00 euros with a fast model; "Fast model usage limit", no number given | Not stated | Angel 7.25 euros a month plus taxes; prices apply to academic, non profit and personal use | More than 240 million papers from the Semantic Scholar index; citation networks on Free | Chatting with papers and uploaded files |
| Elicit | Unlimited paper search across more than 138 million papers; Research Agent has "Limited usage" | Not stated | Pro $49 a month; Scale $229 a month | Source viewing for answers; Zotero import | Systematic reviews |
| Consensus | Basic paper search; 10 Pro queries and up to 3 Deep reviews a month | Not stated | Pro $12 a month; Deep $45 a month | 400M+ peer reviewed papers and scholarly works | Quick answers from published studies |
| Semantic Scholar | Free, AI driven search and discovery tools | Not stated | Not stated | Over 200 million academic papers | Finding and reading papers |
| ChatGPT | "Unlimited text chats with GPT-6" on Free; uploads, image creation and deep research marked "Limited" | Not stated | Go, Plus and Pro exist; no dollar amount shown on the page as read | Not stated | General science explanations |
| Gemini | Free with a Google Account; Deep Research and image generation listed, no limit given | Google Account | Google AI Pro $19.99 a month | Not stated | Explanations and Google account tools |
| Claude | "Limited usage" every 5 hours and every week; web search included; Research not included | Not stated | Pro $20 a month, or $17 a month billed annually | Answers from the web | Careful, long explanations |
| Whizi | 10 messages a day on Whizi AI, Auto and Claude Haiku, no card | Account needed | Pro $29.99 a month | Web search on by default; PDF uploads | Checking one answer with two models |
Whizi loses on two counts. It has no paper index of its own, so Consensus, Elicit and Semantic Scholar find papers it can't. And Whizi needs an account, while Gemini on its Free plan needs only a Google Account you may already have.
The dedicated tools split by job. scienceOS describes itself as an AI research agent for scientists, and its Free plan answers "based on text" while figures need the paid Angel plan. Consensus gives 10 Pro queries a month before you pay. Elicit is built for systematic reviews, with screening of up to 5,000 papers on Pro. Claude's pricing page marks research as unavailable on Free. For an explanation or a calculation, a chatbot is the better fit.
Can science AI be trusted with citations?
Not without checking. The clearest test is the 2023 study by Walters and Wilder in Scientific Reports, which checked 636 citations in papers written by GPT-3.5 and GPT-4. Their finding, in their words: "55% of the GPT-3.5 citations but just 18% of the GPT-4 citations are fabricated."
Even real citations went wrong. The study found substantive errors in 43% of GPT-3.5's real citations and 24% of GPT-4's, and fewer than 10% carried a link. The paper tested older models without web search, so today's tools do better. The rule still holds: a paper you can't open is a guess. How to fact check AI answers turns that into a two minute routine.
What are the best science AI prompts for studying?
Study prompts work when they say what you already know, so the answer starts at your level. For physics and chemistry problems, physics AI solver and chemistry AI solver go deeper. These six are for any science course.
- Explain at my level: "Explain [concept] to a [grade] student who already understands [prerequisite]. Use one everyday example, then ask me one question to check I got it. Don't answer your own question."
- Check my reasoning: "Here is my explanation of [phenomenon]: [paste]. Find the first claim that is wrong, say why in one sentence, and stop there."
- Misconception hunt: "List the three most common misconceptions students have about [topic], and give one question for each that would expose it."
- Experiment design: "I want to test whether [variable] affects [outcome] using [equipment]. Name the independent, dependent and control variables, then list three things that could ruin the result."
- Quiz me: "I just studied this: [paste notes]. Ask me five questions that need me to apply it, one at a time. Wait for my answer before the next."
- Units audit: "Here is my working: [paste]. List the unit on every line and flag any line where the unit does not match the quantity."
What prompts work for papers, data and diagrams?
Research prompts work when they set the source rule or give the model the document to read. These six cover finding papers, reading them and checking a figure.
- Find papers: "Find five peer reviewed papers on [topic] published after [year]. For each give authors, journal, year, DOI and one sentence on the finding. Mark any you could not confirm exists."
- Summarize a paper: "Here is a paper: [paste the text, or attach the PDF]. Give the question, the method, the sample size and the main result in five lines, quoting the sentence for each. Say what the authors list as limits."
- Read the statistics: "In this paper, what is the sample size, what test was used and what does the p value mean here? Explain it in plain words and note anything that makes the result weaker than it sounds."
- Compare two studies: "Compare these two papers on [topic]: [paste both abstracts]. Make a table of method, sample, result and limit. Tell me which finding I should trust more and why."
- Read a diagram: "I've attached a figure from a paper. Describe it in words first: axes, units, groups and any error bars. Tell me anything unclear before you say what it shows."
- Explain a result to a non scientist: "Summarize this finding for a general reader in 80 words: [paste abstract]. Don't say it proves anything, and say who the study did not cover."
Prompt 8 matters most. Asking for the sentence behind each line lets you find a misreading in seconds. See summarize PDF with AI for long papers.
What mistakes ruin a science AI answer?
Four mistakes cause most bad answers, and each has a quick fix.
- Citing a paper you never opened. The fix: search its title, open it and find the result yourself.
- Asking a chatbot for papers with web search off. The fix: switch search on, or use a paper tool like Consensus or Semantic Scholar for the list.
- Dropping units halfway through a calculation. The fix: ask for units on every line, as in prompt 6.
- Treating an abstract as the whole study. The fix: ask for the sample size and the limits, as in prompt 8, and read the methods yourself.
A fifth is the quiet one: no level. A model that doesn't know your course answers for nobody in particular.
Where does Whizi fit as a science AI?
Whizi is one app for science questions from Claude, GPT and Gemini instead of a pile of single purpose tools. It suits you if you want a second opinion on an answer you plan to rely on.
Web search is on by default in each chat, and you can upload a paper as a PDF, Word or Excel file. Scanned PDFs get a vision transcription pass, so a photocopied article is readable. You can also attach an image of a diagram to a chat message for the model to look at, so prompt 11 works as written.
The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. On Pro, side by side Compare runs two models on one prompt, so you read only the lines where they disagree. See Whizi pricing for what each plan holds. A paper index like Consensus or Semantic Scholar is still the faster way to find papers.
What can a science AI not do?
It can't promise a correct answer, and it can't replace the paper itself. It predicts likely text, and a likely sentence isn't always a true one. Why AI gets things wrong explains the mechanism.
It also can't run your experiment, review your methods or stand in for peer review. The free tiers are thin as well. scienceOS says its Free plan answers from text only, and Claude's Free plan lists no research feature. For medical, safety or lab decisions, a qualified person has to sign off. Related guides: AI research assistant and biology AI.
- Decide whether you need an explanation, a calculation or a list of papers
- Switch web search on and name the source types
- Search the title of every paper the model names
- Keep units on every line of a calculation
- Ask a second model when the answer matters
Frequently asked questions
Is there a free science AI for finding papers?
Yes. Semantic Scholar is free and covers over 200 million papers, and Consensus lists 10 Pro queries a month on its free plan, as of November 2026. Elicit lists unlimited paper search. None of the three pages gives an accuracy score, so open each paper.
Can ChatGPT answer science questions?
It can explain most school and undergraduate topics, but it makes mistakes on details. ChatGPT's Free plan lists unlimited text chats on GPT-6, as of November 2026. Ask for sources, then open them, because one 2023 study found 55% of GPT-3.5's citations were fabricated.
Can a science AI read a PDF of a paper?
Yes. scienceOS accepts PDF, Word, PowerPoint and Excel files, with 128 pages per document on Free. Chatbots read papers too. Paste the text or attach the file, tell the model to quote the sentence behind each claim and check those quotes.