That question lies at the core of research by Dr. Yifei Wang, Dr. Ashkan Eshghi, Dr. Yi Ding and Prof. Ram Gopal at Warwick Business School. Their study explores a subtle consequence of large language models, such as ChatGPT: when AI rephrases human writing, it may soften the emotional signals contained in the original. In an age when online language is routinely mined to understand public opinion, that small shift could have surprisingly large consequences.

The emotional character of what people write online can help researchers, businesses and policymakers gauge attitudes toward products, events and public issues. A flood of angry posts may suggest one kind of public response, while enthusiastic comments suggest another. If AI increasingly sits between what people initially write and what they finally publish, however, the resulting text may no longer be a perfectly faithful record of those feelings.

To investigate, the researchers turned to a vast collection of tweets about climate change written between 2019 and 2021, before generative AI became an everyday writing assistant. They selected 50,000 tweets and asked a language model to paraphrase them while keeping within Twitter’s character limit. They then compared the emotional tone of the rewritten versions with that of the human originals.

The pattern was striking. Rephrased tweets tended to move toward emotional neutrality. Strongly positive language became less positive, while strongly negative language became less negative. The AI model appeared to compress emotional extremes, producing language that was more restrained than the material it had been given. Negative feeling was dampened somewhat more strongly than positive feeling, and the rewritten tweets showed a small average movement toward greater positivity overall.

This effect persisted even when the researchers changed their instructions. Explicitly telling the model to preserve sentiment did not eliminate the tendency. Nor did focusing the instruction on grammar correction. Testing a more advanced language model produced the same broad pattern, in that the emotional smoothing remained.

That matters because researchers increasingly treat digital conversation as a window into human attitudes. Imagine a thermometer that consistently nudged very hot and very cold readings toward room temperature. It might still detect broad changes, but anyone relying on its measurements could misunderstand the intensity of what was happening. AI-assisted writing may create a similar problem for sentiment analysis. The words remain readable and plausible, but the emotional temperature can shift.

While this phenomenon has implications for researchers assessing contentious digital discourse on political or environmental issues, for example, it can also affect the commercial sphere. The team examined thousands of product reviews and again found that AI-rephrased text showed reduced positive and negative sentiment. That suggests the phenomenon can travel across different kinds of online writing, potentially affecting both studies of public opinion and assessments of customer satisfaction and reputation.

The implications became clearer when the researchers assessed what happens when these softened messages are used to draw scientific conclusions. The team recreated an earlier study examining how the COVID-19 pandemic affected sentiment in climate change tweets. They first ran the analysis using the original human-written messages, then repeated it after replacing them with AI-rephrased versions.

The broad direction of several relationships remained intact, but their apparent strength changed. Some effects became weaker, while the statistical importance of certain factors shifted. In other words, the systematic AI-mediated alterations accumulated across thousands of messages until they were capable of changing what researchers might conclude from the data.

As AI writing assistance becomes ordinary, online language may increasingly represent a mixture of human expression and machine refinement. Researchers studying that language could therefore face a difficult problem. Are they measuring what people actually felt, or what people felt after an algorithm made their words sound calmer, cleaner and more conventional?

Crucially, the study proposes ways to recover some of the emotional information that AI editing can blur. The first approach treats the model’s influence as a pattern that can be learned. Because the emotional changes produced by rephrasing were sufficiently systematic, the researchers trained predictive models to estimate the sentiment of the original human text from the altered version. Techniques such as linear regression, neural network regression and random forest regression all performed effectively in their tests.

That approach works best when we already know that a piece of writing has been rephrased by AI. Real life is messier. A researcher encountering a post online may have no way of knowing whether it was written exclusively by a person, polished by a language model or produced through some combination of the two.

So, the team developed a second strategy. They fine-tuned language models to reverse the effects of AI rephrasing and produce text whose sentiment more closely resembled the original human writing. One fine-tuned model narrowed the emotional gap. It also showed an ability to distinguish between material needing correction and untouched human text. When tested on 10,000 AI-rephrased tweets, it modified every one. Among 10,000 original tweets, it changed only 38.

The research does not suggest that people should necessarily stop using AI to improve their writing. These tools can make communication clearer, faster and more accessible. Rather, the research reveals a hidden trade-off: a sentence can become more polished while becoming slightly less representative of the feeling that prompted it.

This phenomenon will grow more important as AI-assisted language fills social media, reviews, forums and other spaces. Public sentiment has value precisely because it captures the uneven, emotional texture of human response. Anger, excitement, anxiety and enthusiasm are part of what the data tells us.

Dr. Wang and colleagues therefore point toward a challenge. As machines help us to express ourselves, preserving authenticity may require us to understand what those machines change. The future of authentic communication may depend on ensuring that when AI improves our words, it preserves our distinctly human feelings.