Co-writing with AI boosts fluency, yet compromises originality. Here are arguments to design for accountability, originality, and reader trust.

ChatGPT makes us faster and better writers on the surface, but it also makes us more predictable and dull. That pull is not only a stylistic quirk, it is a big cultural shift, with consequences to be understood—and here I’m watching out for using em dashes, which were nearly canceled.

Why this, why now?

When millions of creators rely on similar assistants, we risk saturating feeds with fluent, perishable prose and images that feel instantly familiar and just as instantly forgettable. The design problem is therefore twofold: keep a human clearly accountable for what gets said and shipped, and keep enough friction in the process to preserve strangeness, stance, and the kind of originality that can’t be averaged away (European Union 2024; European Union 2016).

HCI studies of co-writing—from open datasets to in-the-wild tools—have moved beyond demos into careful observation of real practice (Lee, Liang, and Yang 2022; Yuan et al. 2022). Large field experiments in workplaces show AI assistance lifting average performance and compressing variance (Brynjolfsson, Li, and Raymond 2023).

In work communication, adherence to a shared vocabulary is paramount to good understanding and communication. It is informative communication, in opposition to phatic information (used for social interactions and emotions). And AI does not only use informative, but also phatic communication. It creates concise, information-packed messages, with a soft delivery.

But AI also helps to litter the internet. Platforms are tightening against scaled, low-value content: Google’s March 2024 core update targets “unhelpful” mass production, and YouTube requires disclosure for realistic synthetic media (Google Search Central 2024; YouTube 2024). If you care about voice and trust, now is the moment to design workflows that prevent homogenization and make a human signer visible.

The solitary author is dead. The ensemble writing is rising.

The Wordcraft project, which set professional authors loose in a language-model editor, reports a familiar blend of delight and friction: the system is astonishing at rephrasing, proposing plot turns, and keeping momentum, yet it leans into formula unless the human actively steers tone and intent (Yuan et al. 2022).

Good for divergence, bad at convergence✖AI cross-pollinate ideas with more power than humans can, but it tends to mediocrity when it comes to defining the core of the work—formulaic, unless told to do differently.

CoAuthor, a CHI dataset of 1,445 GPT-assisted sessions, shows that models are productive collaborators for ideation and revision but that the value is highly sensitive to interface pacing and framing—an interaction design problem as much as a model-quality one (Lee, Liang, and Yang 2022).

If the machine accelerates the middle of the draft, the human must insist on defining the beginning and end—what the piece is for, what it claims, and what it risks. The machine is a good autopilot, but the human is still deciding where to go through the more or the less travelled paths.

The problem of accountability

A model can draft a policy; it cannot sign it. In Europe, this is not just culture, but law. The EU AI Act obliges human oversight for high-risk systems and gives differentiated duties to providers and deployers, making clear that someone—not something—remains answerable (European Union 2024).

Separately, GDPR Article 22 limits solely automated decisions with legal or similarly significant effects and guarantees a right to human review (European Union 2016). Treat these as design constraints, not compliance chores: every AI-shaped artifact should have a visible human signer who understands what changed and why, and who can respond to errors and harms. This discipline is not anti-automation; it is the minimal condition for public trust.

“Delving into” the problem of mediocrity

Alignment methods such as reinforcement learning from human feedback make models more generally helpful and safer across contexts, but they also have aesthetic consequences.

An ICLR 2024 paper finds RLHF (Reinforcement Learning from Human Feedback, a machine learning technique) reduces output diversity both per input and across inputs relative to supervised fine-tuning—a formal account of why aligned models feel “balanced” (Kirk et al. 2024).

Creative ideation experiments likewise show that groups feel more creative with LLMs while their outputs converge in structure and theme (Lee, Liang, and Yang 2022).

At workplaces, AI help lifts quality and compresses dispersion excellent for service consistency, less so for stylistic differentiation (Brynjolfsson, Li, and Raymond 2023).

This is what it means to operationalize central tendency: next-token probabilities become next-idea probabilities, and unless we deliberately perturb the process, the “nuanced” median becomes the default.

There is a systemic worry: if models are trained increasingly on model-generated data, tails of the distribution can vanish—that is, the very different input provided by us, humans, at our best and worse creative outputs. This is a dynamic dubbed “model collapse” (Shumailov et al. 2023).

Warning signs

While the ability of AI to create 101 creative ideas for a new project, teams may tend to fallback in known, safe spaces and feel more creative than they really are. Relying on this dynamics may prevent teams of going the extra mile and offer innovative input.
At the same time, monotonous work gains speed and precision.
And what happens if there’s shortage of our innovative input to the model?

Flood and fatigue

Everytime the cost of production drops, spam rushes in. Google’s March 2024 update explicitly targeted scaled content abuse and “unhelpful” pages made for search engines rather than humans (Google Search Central 2024).

YouTube now asks creators to disclose realistic synthetic media and is tightening monetization rules around repetitive, mass-produced videos (YouTube 2024; YouTube Help 2024).

The message is clear: platforms will ratchet their policies to defend scarce attention. For creators and teams, the incentive is to foreground provenance, voice, and the human signer—traits low-value automation struggles to fake at scale.

What creativity needs from tools (and from us)

Sampling knobs matter—temperature and top-p can increase surprise—but knobs aren’t a strategy (OpenAI 2025). A more robust pattern is process choreography:

Human sets stance and stakes first ✖ Draft the thesis, constraints, and intended reader effect before any autocomplete.
Model expands, varies, and challenges ✖ Use the system to produce far-apart framings and counter-arguments, not just paraphrases. Mandate “5 divergent options” rather than “rewrite clearer.”
Human edits and signs ✖ Accept additions selectively, restore oddities, and take responsibility for claims.

The role humans still can affect✖ Fresh collection of data from the world, the perception of one’s own lived experience, errors, silliness, the art of getting lost, serendipity, happy accidents and a reboot in framing thought.

A theoretical lens

Byung-Chul Han, Umberto Eco

Byung-Chul Han describes a culture of transparency in which the imperative to make everything visible and comparable smooths away friction; trust is replaced by control, and difference by synchronization (Han 2015). Language models—trained on vast corpora and then alignment-tuned to avoid offense and maximize general helpfulness—help realize that cultural program in text: they excel at producing sentences that are instantly legible and broadly acceptable. The risk here is not censorship but thinning. When every line must be immediately digestible, the inassimilable detail—the aside that pricks your attention, the analogy that asks for a pause—loses its habitat.

Umberto Eco’s notion of the open work offers a complementary directive. An open work invites interpretive participation; it composes for a reader-performer who completes the piece in practice (Eco 1989). Co-writing with AI could embody openness if the human author first sets a horizon of meaning—what the piece is trying to do, what questions it wants to leave alive—and then treats the machine as a performer within those coordinates. Too often, though, our tools pre-compose the options, privileging optimized continuations over genuine exploration. The design consequence is straightforward: interfaces should protect zones of ambiguity in which human intent can gestate before the assistant proposes polished alternatives; suggestions should be collapsible by default; and “show me five incommensurate framings” should be a first-class action, not a hack. Openness is not a glut of options; it is a choreography of uncertainty and choice.

Practical advice for working with AI and content

Name the signer. Every AI-touched document ships with a named human accountable for content and consequences (European Union 2024; European Union 2016).
Stage divergence before convergence. Human thesis → model expansions/counter-frames → human edit and sign-off (Lee, Liang, and Yang 2022).
Constrain the assistant. Disable rewrite-as-you-type for first passes; add a “wild-card” action that forces divergent options (OpenAI 2025).
Log provenance. Track suggestions accepted/rejected; include a short editor’s note when stakes are high.
Ground in the world. Regularly inject non-synthetic inputs—interviews, observations, quotes—so your model is echoing your reality, not the web’s median (Yuan et al. 2022).

Problems and Possibilities in co-writing with AI