Adobe’s approach to generative AI in digital documents
Over more than decade, Adobe has delivered hundreds of intelligent capabilities across our creative, experience and document applications, providing tremendous value across billions of Adobe Acrobat subscribers worldwide.
Generative AI brings even more exciting possibilities — and some new challenges. Starting with Adobe Firefly, our family of generative creative models, we’ve made a commitment to thoughtful AI innovation and leadership. We continue to demonstrate this commitment in how we develop and deploy our technologies, in the practices we develop and through collaborations with industry and government partners to help create a world where game-changing discoveries can flourish alongside responsibility.
Today we’re embarking on our next chapter of generative AI innovation by introducing our first capabilities for digital documents. AI Assistant in Reader and Acrobat, now in beta, is a new AI-powered conversational engine deeply integrated into PDF workflows.
We are taking the following steps to put the right capabilities and guardrails in place so our customers — from individuals to the largest enterprises — can use AI Assistant with confidence.
Building trust
Generative AI is an incredibly exciting technology that's already delivering tremendous value to our customers. Generative AI is also still in its early stages and improving the technology is an ongoing journey. To help build trust with customers, Adobe is taking a multilayered approach:
- AI ethics testing and reviews: All AI Assistant features in Reader and Acrobat, including third-party LLM integrations, go through Adobe’s responsible AI Ethics governance process and are developed and deployed in alignment with the company’s AI Ethics principles of accountability, responsibility, and transparency.
- Guardrails for LLMs: When working with third-party LLMs, Adobe contractually obligates them to employ confidentiality and security protocols that match our own standards. We specifically prohibit third parties from manually reviewing or training their LLMs on Adobe customer data. We also include robust testing and evaluation methodologies in pre- and post-processing and engineering processes.
- Data security protocols: AI Assistant features in Reader and Acrobat are governed by data security protocols, including testing and evaluation methodologies in pre- and post-processing, and evaluation, testing and evaluation built into our engineering processes.
- Keeping humans in the loop: AI Assistant includes attributions in generated responses, making it simple for customers to confirm where the information came from; an in-app message reminds customers to double-check the source of the answers AI Assistant provides.
- Built for business: AI Assistant includes enterprise-grade data security and information governance.
- Ongoing customer feedback: Generative AI is still emerging and improving the technology is an ongoing journey. Adobe provides customers with multiple channels — from in-app feedback and advisory boards to community groups and social listening programs — and encourages active conversations to help the company identify and address issues quickly.
Leveraging large language models (LLMs)
As a leader and visionary in artificial intelligence, Adobe builds foundation models in the categories where we have deep domain expertise and partners with best-in-class large language models (LLMs) for text-based experiences.
We take an agnostic approach to LLM-integrations, curating the best technologies, partners and models to deliver the right output for our customers’ needs. For AI Assistant features in Reader and Acrobat, we are currently leveraging the Microsoft Azure OpenAI Service. We will continue to explore and test a variety of technologies to provide customers with quality, responsible experiences.
When working with third-party LLMs, Adobe contractually obligates them to employ confidentiality and security protocols that match our own standards. We also evaluate all LLM integrations in accordance with Adobe’s AI Ethics process and principles. We specifically prohibit third parties from manually reviewing or training their LLMs on Adobe customer data. We also include robust testing and monitoring methodologies in pre- and post-processing and engineering processes.
Liquid Mode: A building block for document intelligence
Liquid Mode — Adobe’s breakthrough reading experience that delivers an easier way to read documents on mobile — is powered by proprietary AI models that offer a deep understanding of PDF structure and content. Liquid Mode turns static information into dynamic data, automatically reflowing PDF documents to make them more readable and accessible on any sized screen. On the 30th anniversary of Adobe Acrobat and PDF, Liquid Mode achieved a major milestone, with customers using the feature to read more than 1 billion files on mobile devices. The technology has revolutionized reading experiences for individuals of all abilities and was recognized as one of Time Magazine’s 2023’s Best Inventions.
At the heart of Liquid Mode is the proprietary AI and ML-powered models that identify the elements of a PDF — like headings, paragraphs, images, lists, tables and more — and understand the hierarchy and ordering, making it possible to reformat a static PDF into a more dynamic, customizable experience.
Using Liquid Mode, customers can create intelligent outlines, collapse and expand sections and search text for faster navigation. They can also adjust things like font size and spacing between words, characters and lines to suit their reading preferences and abilities. Words also become resizable and reflowable, images tappable and expandable and tables fully responsive.
The experience our engineering and design teams gained by building Liquid Mode and then learning how customers get value from it is foundational to what we’ve delivered in AI Assistant. The ability to deeply understand PDFs enables AI Assistant to analyze and surface insights from digital documents with greater precision now and will allow us to continue to refine the outputs even further into the future.
Into the future
We’re proud of Adobe’s role in transforming the definition of PDF and the value we’ve delivered over the past 30 years. We believe the next 30 years will be defined by responsible innovation and look forward to working with our customers, our industry and the broader community to imagine and invent new ways to learn, create and collaborate with digital documents across our personal and professional lives.
Forward-Looking Statements
This blog contains “forward-looking statements” within the meaning of applicable securities laws, including those related to Adobe’s expectations, plans and vision for future generative AI capabilities and offerings and the expected benefits to Adobe. All such forward-looking statements are based on information available to us as of the date of this blog and involve risks and uncertainties that could cause actual results to differ materially. Factors that might cause or contribute to such differences include, but are not limited to: failure to innovate effectively and meet customer needs; issues relating to the development and use of AI; failure to realize the anticipated benefits of investments or acquisitions; failure to compete effectively; damage to our reputation or brands; service interruptions or failures in information technology systems by us or third parties; security incidents; failure to effectively develop, manage and maintain critical third-party business relationships; risks associated with being a multinational corporation and adverse macroeconomic conditions; failure to recruit and retain key personnel; complex sales cycles; changes in, and compliance with, global laws and regulations, including those related to information security and privacy; failure to protect our intellectual property; litigation, regulatory inquiries and intellectual property infringement claims; changes in tax regulations; complex government procurement processes; risks related to fluctuations in or the timing of revenue recognition from our subscription offerings; fluctuations in foreign currency exchange rates; impairment charges; our existing and future debt obligations; catastrophic events; and fluctuations in our stock price. For a discussion of these and other risks and uncertainties, please refer to Adobe’s most recently filed Annual Report on Form 10-K and other filings we make with the Securities and Exchange Commission from time to time. Adobe undertakes no obligation, and does not intend, to update the forward-looking statements, except as required by law.