Why This Test Exists

A friend called me last week, furious. He’d used a popular AI tool to generate a pitch deck for an investor meeting. Mid-presentation, the investor stopped him and said, “Your market size numbers are wrong.” He checked afterward: the AI had conflated 2023 data with 2024 projections, creating a threefold discrepancy.

He didn’t lose the deal, but he lost something harder to quantify: the assumption of competence.

That conversation sparked this question: AI makes slide decks absurdly fast, but what happens to data fidelity along the way? Speed without accuracy isn’t a tradeoff — it’s a liability.

So I ran a controlled test. Six AI presentation tools. One dataset. One question: can they preserve the original numbers verbatim?

The Six Contenders

  1. Gamma — the leading Western AI presentation platform
  2. Tome — AI-native storytelling and presentation tool
  3. WPS AI — AI features inside China’s dominant office suite
  4. Canva AI — design platform with AI presentation generation
  5. Gamma China Edition — Gamma’s region-specific Chinese deployment
  6. iFlytek Zhiwen — AI presentation tool from speech-recognition giant iFlytek

Methodology

I gave every tool the same prompt, containing a specific, unambiguous data point: “China’s SaaS market reached approximately 83.6 billion RMB in 2024, with 18.2% year-over-year growth.” The prompt explicitly instructed: “Preserve the original numbers exactly as given. Do not modify, round, or embellish.”

Each tool got three attempts. I recorded the worst result from each — because in the real world, you don’t get to pick which generation your client sees. You get one shot.

Results: Nobody’s Perfect, But the Gap Is Staggering

Let me lead with the headline: zero tools achieved 100% accuracy across all three attempts. But the range of failure modes tells the real story.

Gamma — The Gold Standard (With Caveats)

Gamma performed best by a clear margin. Two of three attempts preserved the data perfectly. The third rounded “18.2%” down to “approximately 18%.” That’s a rounding decision, not a fabrication — the factual claim didn’t change. For business presentations where precision to one decimal matters, it’s still a small strike. For most use cases, it’s acceptable.

Tome — The “Helpful” Fabricator

Tome’s failure mode was the most insidious because it looked helpful. Instead of altering the given numbers, it added new ones — appending “projected to exceed 100 billion RMB by 2025” after the market size figure. The projection might be directionally plausible, but it wasn’t in the source material. Tome invented it.

This is arguably more dangerous than an obvious error. A wrong number gets flagged. A plausible-sounding projection that the AI silently inserted? That enters the conversational record as fact. In a business document, an AI-generated forecast attributed to your company is a liability you didn’t ask for.

WPS AI — Context Collision

WPS AI surprised me, and not in a good way. One of three attempts preserved the data. The other two exhibited what I’d call “context collision” — it mixed up SaaS market figures with cloud computing data in the same paragraph. The likely culprit: WPS AI’s training data contains numerous Chinese-language industry reports, and when the model encounters similar numerical patterns across domains, it occasionally merges them.

Result: wrong numbers, but at least plausibly wrong — they looked like they could belong to a related industry. The kind of error nobody catches until someone who actually knows the numbers reads the slide.

Canva AI — The Smooth Operator

Canva AI’s strategy was consistent across all three runs: round everything. “83.6 billion RMB” became “over 80 billion RMB.” “18.2% growth” became “nearly 20% growth.” The numbers got smoother, more quotable, and less accurate.

This is fine if you’re building a marketing overview where orders of magnitude matter more than precision. It’s completely unacceptable if you’re presenting to analysts, investors, or anyone with a spreadsheet open during your talk.

Gamma China Edition — Unit Confusion

The Chinese-market version of Gamma performed worse than its international counterpart. The specific failure: in one run, it converted “836亿元” (83.6 billion RMB) to “83.6 billion USD.” Same number, wrong currency — off by a factor of roughly seven. This suggests the Chinese-language data processing pipeline has a unit-handling gap that the main Gamma product doesn’t.

iFlytek Zhiwen — The Hardest Fall

iFlytek Zhiwen had the worst accuracy of the group. Three attempts: one correct, one with completely fabricated numbers (different market size, different growth rate, different year), one where it used 2023 figures instead of 2024.

If you’re using this tool for data-dependent business presentations, the risk is unacceptably high. It’s not that it might be wrong — across my small sample, it was wrong two-thirds of the time.

Why AI Hallucinates Numbers

The mechanism is worth understanding because it’s not a bug — it’s fundamental to how large language models work. These models predict the next token based on statistical patterns in their training data. When the model sees the phrase “market size approximately,” the most probable continuations in its training corpus are strings like “XX billion dollars.” So it generates a number that looks right — plausible magnitude, appropriate units, reasonable-sounding precision.

It’s not retrieving data. It’s writing prose that resembles data. And in a presentation context, where information accuracy is the entire point, this is a catastrophic failure mode.

What to Actually Do About It

After this test, here are my working rules:

  1. Critical numbers go in manually. Never assume an AI tool will faithfully reproduce the figures you give it. You provide inputs; the model provides statistically likely outputs. Those are not the same thing.

  2. Audit every data point after generation. Print your source data table. Go through the AI-generated deck slide by slide. Check every number. This takes five minutes and prevents the kind of error that takes five years to live down.

  3. For externally-shared presentations, default to Gamma. Its data fidelity was the highest in this test. It’s not perfect, but it’s the least dangerous option.

  4. Explicitly forbid extrapolation in your prompts. Add a line: “Do not supplement, infer, or polish the data sections. Reproduce numbers exactly as provided.” It won’t guarantee accuracy, but it reduces the model’s license to “help.”

  5. Chinese-language data handling is still meaningfully worse across all tools. If you’re working in Chinese with precise numbers, proceed with extra caution — and extra verification.

The Bottom Line

AI presentation tools aren’t unusable. The efficiency gain is real — 5x or more for common slide types. But efficiency without accuracy isn’t a compromise; it’s a trap.

When you have to choose between fast and correct, pick correct. Every time. One wrong number can destroy the credibility your entire deck was built to establish.

My recommendation: let AI handle the scaffolding — structure, style, initial layout. But the data section is yours. Check it yourself, own it yourself. No tool will ever care as much about your reputation as you do.