Tone Control (www.robinsloan.com)

Writer Craig Mod warns that today's "new systems" produce a smoothed-over, overly familiar English — a single default tone that feels as mismatched as the wrong color temperature on every streetlamp. He imagines a near future where generative tools flatten stylistic diversity so thoroughly that "every written thing will be the wrong tone," and hopes the pendulum might swing back toward crisp formality and clearer tonal control. For the AI/ML community this is a practical design and research problem: model outputs aren’t just about factual correctness but about voice, register, and appropriateness. Solutions include explicit style conditioning (style tokens, prompts, or control codes), fine‑tuning or adapter/LoRA layers for brand or authorial voice, disentangled latent representations for content vs. style, and RLHF/objectives that reward tonal appropriateness. Evaluation needs to move beyond perplexity to metrics and human judgements for tone, cultural fit, and user preference. The risk of homogenization matters for UX, trust, and cultural expression, so researchers and product teams should prioritize controllable generation, richer training data representing diverse registers, and tooling that lets users select or preserve distinct tones rather than defaulting to a single, inoffensive voice.