Most generative AI and custom model projects will be a bust: Gartner

Analyst firm Gartner thinks at least half of all generative AI projects “will overrun their budgeted costs due to poor architectural choices and lack of operational know-how,” and most organizations that try to build custom models “will abandon their efforts due to costs, complexity and technical debt in their deployments.” Those findings headline the Hype Cycle for Generative AI the firm published last week, which considered 30 AI technologies and found none have reached the “plateau of productivity” – Gartner-speak for products and technologies that have gone through two or three generations of evolution, stabilized, and produce verifiable real-world benefits. To reach the plateau, tech ascends the Peak of Inflated Expectations, falls into the Trough of Disillusionment, and slowly climbs a Slope of Enlightenment. Gartner labels Domain-specific GenAI models – models built from scratch or fine-tuned on domain data – as likely to produce superior results and fewer hallucinations compared to the output of general-purpose models in fields such as healthcare, finance, law and other industries. But the firm advises building these models “requires significant compute resources, specialized expertise and ongoing maintenance,” and rates their maturity as “adolescent” and placed it just before the Peak of Inflated Expectations and at least two to five years from becoming mature enough for mainstream use. Just one of the technologies Gartner considered is climbing the slope: Generative-AI-enabled applications such as coding assistants, graphics and video creation, and summarizing content. The firm worries that intellectual property concerns and the tendency to create inaccurate output continue to plague these tools but feels rapid evolution of underlying models means these applications are quite mature – as shown by over half the target market having adopted such applications. The Hype Cycle rates AI agent communication protocols – a specification that defines the rules that allow agents to interact with each other and the wider environment to enable an AI agent to interact with its environment or – as the least mature AI technology. The analyst firm notes that Model Context Protocol (MCP) and agent-to-agent protocol (A2A) are currently the most popular such tools, but says several alternatives are already emerging and that all are evolving quickly as early adopters find weaknesses and omissions. Gartner thinks two technologies – Disinformation Security and World Models – have the greatest potential impact. The analyst firm describes the first as tools that help an organization fight back against disinformation campaigns that use deepfakes, impersonation, and other content to besmirch a company or individual – or to create content involved in cyberattacks. “Attack vectors are varied and can include using GenAI-created content to fool voice or face biometric authentication, or to trick identity verification processes used in account recovery workflows,” Gartner’s analysts suggest. “Once authenticated as the perceived user, malicious actors can take nefarious actions such as planting ransomware, stealing intellectual property, theft of funds and spreading disinformation.” The commercialization of open LLMs has been challenging for builders “Most employees have never seen a deepfake of their leadership or someone they know, and this is a liability in the event of an attack against your organization,” the document states. Gartner thinks red-teaming exercises to detect deepfakes are therefore in order and suggests monitoring social media and forums for nasty AI content about your brand. Good luck with that effort, because Gartner rates these tools as five to ten years away from maturity. World Models are abstractions of a physical environment that Gartner says “empower AI to perform more sophisticated prediction and planning tasks, moving beyond mere pattern recognition in observed data. By simulating and understanding the dynamics of environments, AI can better handle uncertainty or missing information and therefore make informed decisions that account for future possibilities and contingencies.” They’re also useful to guide robots through the human world or to create AI-generated videos that more accurately depict the laws of physics. Gartner also thinks that organizations that want to build their AI on open models won’t be able to access the best technology unless they’re willing to consider Chinese tech. “The commercialization of open LLMs has been challenging for builders and many Western tech companies are being selective with releasing open models, which has relegated innovation in this space to China,” the firm observes, adding that while these models “continue to advance in terms of quality and speed of innovation, the innovation ecosystem for open models has shifted east to China.” ®