WRITER Adds Customer-Specific Invisible Watermarks to Palmyra Text

How the watermarks work

When an enterprise enables the feature, text generated by WRITER's Palmyra models will carry a mark tied to that customer's organization rather than to the model provider. WRITER describes text watermarking as a method that subtly steers a model's word choices to create an invisible, machine-readable pattern in the content.

The company argues that common industry approaches fall short. Provider-specific watermarks trace content back to the large language model, which, in WRITER's view, erases the role enterprises and their people play in shaping and aligning the output. WRITER says its platform grounds every AI-generated output in brand voice, customer knowledge, business context and compliance standards, and that its provenance methods should reflect that the content belongs to the customer.

CEO and co-founder May Habib framed the move as a matter of customer choice. "Enterprises don't want their model providers making their AI decisions for them," she said, adding that they want "ownership, control, and configurability."

Tuning detection strength and scope

WRITER says its approach goes beyond switching a watermark on or off. Where regulations allow, customers will be able to tune watermarks to specific use cases and decide what gets marked and how.

A central element is the trade-off between detection strength and content quality. According to WRITER, one-size-fits-all approaches force enterprises to accept a balance that providers set in advance. WRITER instead lets each customer adjust watermark strength within limits it says are designed to protect both sides of that balance.

• A bank's customer disclosures or a pharmaceutical company's regulated communications may call for stronger detection reliability, because tracing the origin of the content matters most.
• A consumer retailer's campaign copy may favor voice and creativity, which could mean more rounds of human review, higher quality output and less detection reliability, since even a few changed words can affect the message.

WRITER also says it will give visibility into how watermarking affects output quality by content type. Where regulations allow, enterprises can choose where to apply it, for example to blog posts but not ad copy, or to email campaigns but not product descriptions.

Campaign tracing and EU defaults

The company's multi-bit approach will let each customer, and even individual teams, embed custom information in the watermarked text, such as a marketing or ad campaign ID. Customers could then trace where that content surfaced and how widely it spread, which WRITER says helps them gauge the reach and effectiveness of a campaign.

For EU customers, text watermarking will be on by default and pre-configured to meet the requirements of the EU AI Act. Non-EU customers can enable it at any time, with controls to tailor it to their business needs.

What it means for enterprise buyers

The announcement lands as disclosure rules for AI-generated content take shape. Roger Beharry Lall, Research Director at IDC, said in the release that as regulators extend disclosure and transparency expectations, enterprises "will want control over how they apply these requirements" and will look for tools that fit the governance frameworks they already use to manage vendors and data.

Habib was sharper in her criticism of other AI labs, calling their approach to watermarking a "hatchet approach" and saying they "checked the compliance box" without considering how customers wanted their content marked. For WRITER customers, the practical change from December 2 is a configurable watermark that identifies the organization behind the text, with EU users covered by default and others able to opt in.