Is AI getting more confident over time?

A hedge word is one of 11 fixed terms that soften a claim: may, might, could, generally, typically, depending on, it depends, consider, some users, your mileage, and not always. The hedge rate is the number of hedge words per 1,000 words of answer text. We counted them straight from the answer text, so no classifier is involved.

On the same 941 prompts, the ChatGPT hedge rate fell from 6.21 in November 2025 to 2.16 in September 2026, a 65.2% drop. The Google AI Mode hedge rate fell from 3.62 to 1.53, a 57.7% drop. Both engines now soften their answers with these words far less often than they did a year ago.

Answers also got shorter. A ChatGPT answer to the same prompts averaged 664 words in November 2025 and 385 words in September 2026. Fewer hedges per word and fewer words together cut the hedge words in a typical ChatGPT answer from 4.12 to 0.83.

Google AI Mode answers were already short and plain. Their hedge words per answer went from 0.72 to 0.5.

The decline was not gradual. On the same prompts, the ChatGPT hedge rate averaged 6.83 from November 14 to 24, 2025. It fell to 3.89 between November 28 and December 9, then to 2.48 from December 11 to 20. The Google AI Mode hedge rate fell from 3.7 to 1.9 in the first week of December.

Each step took two to four days and then held. The recorded model name for these answers was blank in 2025, so the data cannot say whether each step was an engine update or a change in how answers were delivered. Either way, a reader of AI answers saw the change on those dates.

Comparing the last quarter of 2025 with the third quarter of 2026 on the same prompts, ChatGPT used "might" 90% less often per 1,000 words, "consider" 58% less often, and "may" 49% less often.

Two terms moved the other way. "Generally" rose 435% and "typically" rose 116%. These words describe what usually happens. They do not express doubt about the answer. ChatGPT moved from words of uncertainty to words of generalization.

We compared the same prompts in the week before and after each recorded model update. We then compared that change with normal week-to-week variation on the same model and with Google AI Mode over the same dates.

The move from gpt-5-2 to gpt-5-3 cut the hedge rate by 33.5% on 2,893 prompts. A smaller ChatGPT model that did not change on that date moved -0.4%. The move from gpt-5-5 to gpt-5-6 cut the hedge rate by 26.9% on 4,387 prompts, while gpt-5-5 moved -1.0% week over week and Google AI Mode moved -0.5%.

We could not isolate the gpt-5-3 to gpt-5-5 update. Collection paused for five weeks around it, and Google AI Mode fell 23.9% over the same gap. The 96.8% rise in ChatGPT therefore includes other changes during those weeks.

Word counts tell only half of the story. A hedged recommendation is a brand statement that a fixed classifier marks as qualified rather than clear. We kept the classifier fixed by comparing only statements it scored in the same week.

On 11,496 prompts asked before and after the August 8, 2026 handover, the share of hedged recommendations rose from 2.98% under gpt-5-5 to 5.93% under gpt-5-6. Re-asking the same prompts on one model moved it -0.03 and +0.79 points, and Google AI Mode moved +0.05 points. In the same answers, "but" and "however" rose 24.7% and "if you" rose 12.0%.

gpt-5-6 stopped saying a product might suit you. It started saying a product suits you if a stated condition holds, but not otherwise. The answer sounds more certain, and the recommendation comes with more conditions.

The new conditions come in a new voice. In the raw answer text, ChatGPT used "I'd shortlist" 0 times per 1,000 answers in January and February 2026 and 139.6 times per 1,000 answers in September and October 2026. First-person picks such as "I'd start with" or "I'd lean toward" went from 0.3 to 334.9. "Caveat" went from 6.0 to 168.0.

Older wording faded. "Budget-friendly" fell from 61.5 to 7.1 uses per 1,000 answers and "best for" fell from 930.7 to 325.4. Google AI Mode adopted one new phrase: "the catch" went from 0.0 to 18.1 uses per 1,000 answers. The fixed classifier saw the same shift. 31 of the 40 most common hedge phrases it extracted from ChatGPT answers never appeared in its early-window sample.

A reservation is a condition, a fallback framing, a budget limit, a risk warning, or a losing comparison attached to a recommended brand. The reservation rate is the share of brand statements with one. The classifier's own level drifted from week to week, so we compared months only inside the same scoring week and on prompts asked in both halves of the year.

Measured that way, the ChatGPT reservation rate in January 2026 sat 7.8 points below June and in August 2026 sat 3.2 points above it, a rise of about 11 points. Google AI Mode moved +2.6 points in January and +1.6 points in August against the same June baseline.

Position is where a brand appears in the answer, with #1 the first brand named. Within the same scoring weeks, the ChatGPT reservation rate for the #1 brand moved from 9.6% to 10.75%, up 1.15 points. For brands named fourth or fifth it moved from 11.74% to 28.21%, up 16.47 points.

Google AI Mode went the other way at the top. Its #1 brand's reservation rate fell 2.72 points.

Each brand statement carries the need it answers. For pricing and contract needs, the ChatGPT reservation rate rose from 23.23% to 53.26%, so more than half of these statements now carry a reservation. General needs rose the least among the large groups.

Google AI Mode did not follow. Its pricing and contract reservation rate moved from 21.79% to 18.3%.

A market is the category a prompt belongs to on the public Parse index. Among the 18 largest markets with at least 200 ChatGPT statements in the early window, the reservation rate rose in all 18. B2B Sales Intelligence & Data Enrichment rose most, from 6.37% to 29.13%.

Google AI Mode rose in 16 of its 20 largest markets. Reddit and Community Marketing Services fell the most there, by 8.32 points.

We compared each brand with all brands scored in the same week, so the brand's change is relative to the market-wide rise. Of 112 brands with at least 300 statements in each half of the year, EveryPlate gained the most reservations relative to that average, 13.24 points, and Chargebee gained 12.26 points.

PlayStation lost the most, 9.26 points relative to the average, followed by Booking.com at 7.78 points. Meal kits, subscription billing tools, and betting apps make up 8 of the 10 rising brands. The falling list is made of large, broad platforms such as Shopify, Canva, and Google Workspace.

When the same 249,946 brand statements were scored twice, the hedged share was 2.28% at the first scoring and 0.82% at the second. Across scoring weeks, the hedged share ranged from 0.23% to 3.61%. A month-by-month chart of classifier labels would show those scoring changes, not engine behavior. That is why every classifier number in this study compares statements scored in the same week, and why the headline uses word counts.

As a second check on the headline, a separate 4% random sample of all answers, not limited to the same prompts, gave a 62.7% drop for ChatGPT (6.016 to 2.242) and a 59.8% drop for Google AI Mode (3.586 to 1.441) from November 2025 to September 2026.

Exclusions: answers under 50 words; URLs before counting; the month name "May" followed by a date. Some ChatGPT answers from July to September 2026 repeated their own closing paragraphs. Including or excluding them changed the August ChatGPT hedge rate by 0.12 per 1,000 words, so the chart includes them. Answers to the same prompt are re-asked daily, and the week-to-week change on one model is reported beside every handover. Hedge words count surface wording. They do not measure whether the hedge was deserved. Where the charts say ChatGPT or Google across the whole year, the engine is ChatGPT and Google AI Overviews through April 2026 and ChatGPT Search and Google AI Mode from May 2026.

AI answers now sound more certain, but they attach more explicit conditions to the brands they recommend, especially the brands named after the first. A brand that ranks second or fourth is now more likely to be recommended only for a stated budget, team size, or use case.

Read the condition AI attaches to your brand, not only your rank. If the condition is wrong, publish the page that answers it: pricing tiers, the team size you serve, the integrations you support. If you track brand sentiment with a classifier, check that the classifier itself did not change before you report a trend. And when a model version changes, compare the same prompts in the week before and after, against the week-to-week noise, before calling a shift real.