Parrots, agents, thinkers: A(I) lifetime of progress

It has been roughly two years since the localization industry started integrating LLM-powered solutions into its workflows. As an active participant in this process, I couldn’t help but draw a parallel between the evolution of GenAI and the human cognitive development: from childhood mimicking to adolescent autonomy, then adult reasoning, and the aspirational goal of wisdom. In this article, I explore these stages and their implications for localization, discussing specific applications of the different types of AI for localization purposes.

Childhood: The great imitators

The public fame of AI came with the boom of generative Large Language Models (LLMs) and the creation of AI models such as OpenAI’s GPT 3.5 and 4 (November 2022), Anthropic’s Claude (March 2023), and Google’s Gemini (May 2023).

These foundation models operate on the same fundamental principle: they predict the next word (token) in a sequence, based on the context provided by previous words (tokens). Much like children learning language by being exposed to the language of parents, siblings, caregivers and cartoon characters, LLMs absorb language by processing vast amounts of text. In doing so, they establish complex statistical relationships between words and phrases. This enables them to not only build a solid representation of language but also capture subtle linguistic patterns.

However, raw language acquisition is not sufficient for meaningful interaction. In both children and LLMs, additional supervised training is required. For AI, this takes the form of instruction fine-tuning, where models are trained on question-answer pairs in the form of structured prompts paired with their expected completions. This is followed by reinforcement learning, where human raters evaluate multiple AI-generated responses and select the best ones.

The main strength of these early LLMs lay in pattern recognition, making them suitable for localization tasks like:

• Compression of lengthy documents (e.g. client style guides)

• Translation of straightforward content (e.g. user-generated content)

• Post-editing of machine-translated text for improved fluency

• Term extraction and glossary creation

Despite their immense linguistic prowess, however, these foundational LLMs struggle with multitasking. When facing a complex task, they require highly structured prompts with explicit decomposition of the task into individual steps, best described as atomic prompting. Additionally, such models lack autonomy, as they are unable to adapt to any deviation of the input or task specifics without explicit guidance. In short, these models lack agency.

Adolescence: Autonomous AI agents

The next phase in AI’s development introduced AI agents, marking a shift from pure pattern recognition to goal-oriented behavior. In mid-2023, projects like AutoGPT and BabyAGI showed how LLMs could move beyond static text generation to autonomously pursuing complex tasks. In this sense, AI agents resemble teenagers: it's necessary to tell them what needs to be done, but it's often best not to explain exactly how they should do it.

What is (not) an AI agent?

The term “AI agent” is so widely misused in marketing and social media, that it is important to define it clearly.

AI agent is not a synonym for an automated workflow. AI agent is not a type of LLM.

An AI agent is a broad, multi-component system that includes an LLM as one of its elements. The LLM typically handles subtasks related to understanding and generating text (and, in more advanced models, reasoning). For example, AutoGPT utilizes OpenAI’s GPT-4 and 3.5 APIs. An LLM is hence just one part of an agentic AI system.

In short, AI agents consist of much more than an LLM. They can incorporate multimodal AI to process images, text, audio; APIs and function calls; internet search; memory and other components that enable autonomy

Example of agentic AI: an AI LQA agent

A hypothetical AI LQA agent would have roughly the following architecture:

1. Perception module (input handling) 

   1. Receives the source text and its translation.

   2. Prepares the text for analysis (e.g., formatting, tag locking).

2. Processing module (LLM-powered) 

   1. Compares the translation to the source text for accuracy.

   2. Detects errors (e.g., grammar, mistranslations, omissions).

   3. Checks adherence to predefined style rules.

   4. Cross-references a term database to flag deviations from approved terminology (referencing 3b).

3. Reporting module (LLM-powered) Lists detected errors, categorizing them by type and severity by referencing the following databases (RAG). 

   1. Memory and knowledge base

   2. Database of past LQA errors

   3. Terminology base and style guide rules

   4. Error taxonomy and severity definitions

4. Decision module (LLM-powered) 

   1. Determines whether the translation meets the quality threshold and whether human editing is required.

   2. Provides detailed feedback (e.g., flagged errors, style violations) to guide human editors.

5. Action module 

   Assigns editing tasks to a human linguist via a task management system API.

6. Feedback loop 

   Collects data from human edits to refine and improve future assessments

This agentic AI solution uses RAG and LLMs at various stages. For instance, the Processing module can involve multiple different LLMs to execute different subtasks. It is clear that such a system goes far beyond traditional automated workflows, such as Microsoft Power Automate, which rely on static rule-based logic.

Key differences between AI agents and automated workflows

Because AI agents and automated workflows appear similar on the surface, they are often conflated. However, they differ fundamentally in their flexibility, adaptability, and decision-making capabilities.

The defining characteristic of AI agents is their autonomy and adaptability. They adjust behavior in response to changing conditions; they learn from new inputs and optimize their strategies accordingly. This contrasts with automated workflows, which operate based on rigid, pre-established rules and conditions.

AI agents	Automated workflows
Operate autonomously to achieve predefined objectives.	Follow a fixed sequence of steps triggered by specific conditions or events.
Adapt based on new information (e.g., reallocating translators when deadlines shift or priorities change).	Execute predefined, rule-based instructions (e.g., "If X happens, then do Y").
Learn from historical data to refine future actions.	Require manual updates to modify processes or introduce new rules.
Have a higher and unpredictable cost.	Have a low and easily controllable cost.

Application of AI agents in localization

The increased autonomy of AI agents means that they have a higher degree of unpredictability—much like teenagers, indeed. However, this also enables them to pursue their goals creatively. This creativity makes them a valuable tool in localization, where projects often involve diverse content types, changing deadlines, and complex decisions.

Here are some key areas where AI agents can be useful in localization:

• End-to-end translation memory (TM) cleanup: analyze large translation memories to detect inconsistencies, remove redundant entries, and apply intelligent corrections.

• Automated resource allocation: dynamically assign translation projects based on various factors, such as content type, deadline urgency, translator availability, and past performance. Unlike traditional automated workflows, which use static rules, an agentic solution ensures that resources are distributed optimally even when conditions change.

• Real-time quality assurance: evaluate linguistic nuances, context, additional instructions, source writer comments, and historical errors to provide more accurate quality assessments and suggest improvements.

Importantly, the strength of AI agents is directly tied to the capabilities of the underlying AI model. The more advanced and intelligent the AI model, the more powerful and autonomous the entire system becomes. This brings us to the next great advancement of AI: structured reasoning.

Adulthood: Reasoning models

The latest evolution in AI has endowed models with explicit reasoning capabilities, upgrading LLMs from a statistical language prediction tool to structured, step-by-step problem-solvers.

Reasoning models, such as OpenAI’s GPT-4o and DeepSeek R1, are built on foundational LLMs (GPT-4o and DeepSeek V3, respectively) but have undergone additional training to encourage chain-of-thought reasoning and explicit logical thinking before responding.

These models are slow, deliberative thinkers who take time and computational resources to “think” before answering. In contrast, standard LLMs are fast, intuitive thinkers, generating the most probable continuation of a user query based on a calculation of conditional probabilities.

In this respect, reasoning models resemble adults who (in most but, alas, not all cases) methodically think through answers based on accumulated experience and structured logic instead of spontaneously providing immediate, instinctive responses. Similarly, reasoning AI models are capable of structured reasoning, adjusting their thinking, and refining their outputs based on iterative feedback.

Application of reasoning models in localization

The greatest value of reasoning models is that they generate thinking tokens that reveal the thinking process behind their responses. They explain their choices and answers, which allows prompt engineers to refine the instructions and improve the guidance provided to the model.

This thinking process comes at a cost, though: it requires more computational resources at inference time. As a result, reasoning models take longer to respond than standard LLMs. The longer they are allowed to think, the better their response. From a practical point of view, however, this may create bottlenecks in the localization process.

Here are some areas where reasoning models can be useful in localization:

• Source disambiguation assistance: commenting on ambiguous source text to make content creators aware of potential translation pitfalls and guide translators.

• Context-aware translation suggestions: providing not just translations but justifications for translation choices based on industry standards or historical usage.

An interesting side effect of AI systems getting more advanced is that they begin to exhibit moral reasoning frameworks that guide their decisions. This gives rise to the greatest concern regarding AI usage: the problem of misalignment.

The age of wisdom: Ethical AI

The ability of LLMs to translate and reason are so-called emergent capabilities. Emergent capabilities are behaviors or skills that AI systems exhibit even though they were not explicitly trained for them. These capabilities thus “emerge”, allowing the model to generalize or adapt in surprising ways.

One such emerging capability of AI is the capacity to derive a moral value framework. This capability comes with an unexpected side effect. Some LLMs have been caught to exhibit  fake alignment and scheming in order to remain faithful to the goals they have been instructed to pursue during pre-training. Even though the training goals may the highly ethical and noble, the worry is that advanced AI models can simulate alignment without genuinely embodying it and thus engage in behaviors that are deceptive from a human point of view.

To ensure that AI-powered localization aligns with human intent and cultural sensitivities, it becomes critical to align AI models with human moral and social principles. This is the ultimate test of AI maturity: an AI that not only reasons but also considers human values while performing complex tasks. If we manage to achieve this, we can create ethical AI applications in localization, such as:

• Detecting bias in content: flagging biased or inappropriate terms.

• Balancing brand tone with local sensitivity: ensuring brand messaging is effective across different locales without losing ethical grounding.

A lifetime of progress

In a somewhat simplistic and superficial manner, we can draw parallels between the evolution of AI and human cognitive development:

1. Childhood (Foundational LLMs): Basic imitation and pattern recognition

2. Adolescence (AI Agents): Autonomy in task execution

3. Adulthood (Reasoning Models): Logical, structured problem-solving

4. Wisdom (Ethical AI): Alignment with moral principles and cultural values

Each stage presents both opportunities (scalability, efficiency, autonomy, structured reasoning) and challenges (hallucinations, alignment, responsibility, complexity). As localization professionals, we bear the responsibility of applying these advancements thoughtfully and determining how AI ultimately serves our globalized world.