A translation company's eight-year analysis suggests AI could reach human-level language skills by the end of the 2020s, but experts warn that perfecting one skill doesn't necessarily mean we're approaching true artificial general intelligence. Translated, a Rome-based firm, has been measuring how quickly human editors need to fix AI-generated translations compared with human-generated ones. The metric, called Time to Edit (TTE), offers a concrete way to track progress toward what researchers call the technological singularity, the theoretical moment when AI exceeds human control and rapidly transforms society.

How Did Researchers Measure Progress Toward Singularity?

Translated's approach is straightforward but revealing. The company tracked how long professional human editors needed to correct AI-generated translations versus translations made by other humans. In 2015, editors spent roughly 3.5 seconds per word fixing machine-translated text. By 2022, that number had dropped to about 2 seconds per word. For comparison, human editors typically need about 1 second per word to fix another human's work.

Over an eight-year period spanning 2014 to 2022, Translated analyzed more than 2 billion post-edits to build this trend. The consistency of improvement led the company to make a bold prediction: if the curve continued, machine translation could reach human-level editing effort by the end of the decade, or possibly sooner.

"The change is so small that every single day you don't perceive it, but when you see progress across 10 years, that is impressive. This is the first time ever that someone in the field of artificial intelligence did a prediction of the speed to singularity," said Marco Trombetti, CEO of Translated.

Marco Trombetti, CEO at Translated

Why Language Matters for Measuring AI Intelligence?

Language is one of the most difficult challenges in artificial intelligence research. Unlike narrow tasks like playing chess or recognizing images, language requires understanding context, nuance, idiom, and cultural meaning. If an AI system could close the gap between machine and human translation performance, it might signal progress toward Artificial General Intelligence (AGI), the hypothetical stage where AI matches or exceeds human cognitive abilities across a broad range of tasks.

However, the field has moved beyond single-metric forecasts. In March 2026, Google DeepMind published a framework arguing that AGI progress should be measured across multiple dimensions, not just translation. The company identified key abilities that matter for true intelligence:

• Perception: The ability to understand and interpret sensory information from the environment
• Learning: The capacity to improve performance through experience and exposure to new data
• Memory: The ability to store and retrieve information over time
• Reasoning: The capacity to draw logical conclusions from available information
• Executive Function: The ability to plan, organize, and execute complex tasks
• Problem Solving: The capacity to find solutions to novel challenges
• Social Cognition: The ability to understand and interact with other minds

Researchers have also introduced new benchmarks to test frontier AI systems. In March 2026, the ARC Prize Foundation introduced ARC-AGI-3, a benchmark designed to test interactive, experience-based reasoning. Can an agent explore an environment, infer goals, build a world model, and keep learning over time? Humans solve 100 percent of their environments. As of March 2026, frontier AI systems scored below 1 percent on this test.

Another benchmark, Humanity's Last Exam, consists of 2,500 questions built because older tests had become too easy for advanced AI models. The official leaderboard shows a significant gap between systems: Gemini 3 Pro scored 38.3 percent, GPT-5 scored 25.3 percent, o1 scored 8.0 percent, and GPT-4o scored 2.7 percent.

What Does the Real-World Impact Look Like?

Despite optimistic predictions about translation progress, the actual state of AI translation in professional settings tells a more cautious story. According to the European Language Industry Survey 2026, which gathered responses from 1,058 participants across 45 countries, 63 percent of independent translators reported using automated translation in some form. However, only 23 percent of independent professionals rated machine translation or AI quality as high to very high, down from 40 percent in 2025.

The sustainability of translation work has also declined. Only 41 percent of independent translators considered freelancing sustainable in 2026, down from 64 percent in 2023. This suggests that while AI translation is improving, it has not yet reached the point where it can fully replace human judgment and expertise.

Translated itself acknowledged these limitations in its March 2026 assessment. The company stated that machine translation can handle more work but is "not sophisticated enough to remove the need for human judgment" in enterprise localization. The company also forecasts "steady improvement rather than sudden transformation," suggesting that the path to singularity, if it exists, will be gradual rather than abrupt.

In a March 24, 2026 announcement for its Imminent report, Translated argued that AI is moving toward systems that learn through real-world interaction. However, the company also noted that scale alone is showing limits without added capabilities such as reasoning and web interaction.

What Do Experts Say About Singularity Predictions?

The 2026 International AI Safety Report, led by Yoshua Bengio with input from more than 100 experts, offers a measured perspective on AI progress. The report notes that general-purpose AI has continued to improve, especially in math, coding, and autonomous operation. However, it also highlights a critical limitation: models can look like geniuses on difficult tasks, then stumble on simpler ones. Progress could slow, continue, or accelerate, making long-term predictions inherently uncertain.

Translated's translation-based approach to measuring singularity runs into similar problems that plague AGI forecasting more broadly. While perfecting human speech translation is certainly a frontier in AI research, the impressive skill doesn't necessarily make a machine intelligent. Many researchers don't even agree on what "intelligence" means, making it difficult to establish a universal metric for singularity.

Whether hyper-accurate translators are harbingers of transformative technological change or simply impressive engineering feats remains an open question. What is clear is that AI translation has become a significant tool in professional settings, even if it has not yet reached the level of human expertise. The gap between machine and human performance continues to narrow, but the path from narrow translation skill to general artificial intelligence remains uncertain and contested among experts.