Semantic Communications: Sending Meaning Instead of Bits

Shannon deliberately set meaning aside. His 1948 framing treated the semantic aspects of communication as irrelevant to the engineering problem — and that abstraction was so productive that it defined the field for seventy-five years.

Semantic communication asks what happens if you put meaning back in. Instead of reproducing a transmitted bit sequence as faithfully as the channel allows, the system transmits whatever the receiver needs to accomplish a specific task — and success is measured by whether the task succeeded, not by bit error rate.

---

From bit-accurate to meaning-accurate

A conventional stack does three things: source coding removes statistical redundancy, channel coding protects the resulting bits, and the receiver reconstructs the payload as faithfully as the link permits. The metrics are familiar — BER, BLER, throughput, latency, packet loss.

Semantic systems add a layer of abstraction. The transmitter identifies features, intent, or state relevant to a declared task and sends a compact representation. The receiver combines that with context and a semantic model to produce the required interpretation. The quality measure becomes task success, detection probability, control cost, or semantic similarity.

The roadside camera example makes it concrete. Shipping every pixel to an edge server is defensible for archival video. For a collision-warning function, it may be far better to send object trajectories, uncertainty bounds, and a small region of interest. The link then carries a decision-relevant description of the scene rather than an indiscriminate pixel stream.

This is not a licence to discard information casually. A semantic encoder must know the task, the acceptable loss, and the consequences of an incorrect inference. In safety-critical cases, raw data or a verifiable fallback path may still be necessary.

---

How encoding and decoding work

A semantic encoder has three jobs. It observes a source — text, video, sensor data, machine state. It maps that observation into a latent representation or structured message. And it decides what to actually transmit, based on the receiver's available context, the link condition, and the task deadline.

The receiver performs the inverse operation imperfectly by design: it combines the received representation with a model of the source and its own context to infer the relevant meaning. For natural-language traffic, that might mean preserving intent and key facts while changing wording. For sensing traffic, estimating object class, location, and motion from a compressed feature vector.

Shared knowledge is the pivot

The shared-knowledge block is what makes the whole thing work. A sender and receiver using compatible world models can communicate efficiently precisely because they don't need to retransmit what both already know.

That same dependency is the central vulnerability. Model drift, mismatched model versions, or unfamiliar operating conditions turn compression into a wrong conclusion — and unlike a bit error, a wrong inference doesn't announce itself. The receiver produces a confident, plausible, incorrect answer.

---

Where the efficiency actually comes from

The opportunity isn't universal compression. It's task-selective traffic reduction in paths constrained by bandwidth, energy, or latency: edge inference over a congested uplink, swarms of low-power sensors, connected mobility, immersive applications with fast-changing context.

Efficiency arises in several ways:

• Suppressing features that cannot change the decision
• Sending updates relative to the receiver's current scene or knowledge base
• Allocating stronger protection to high-value semantic elements rather than treating every bit equally
• Adapting the representation to available radio resources and the deadline

That third one is worth dwelling on. Conventional channel coding protects all bits equally. Semantic systems can grade protection by importance — an object's classification might warrant far more redundancy than its bounding-box precision. Unequal error protection isn't new, but semantic awareness gives it a principled basis for deciding what matters.

The comparison has to be end-to-end. A system that sends fewer radio bits but adds excessive model computation, synchronisation overhead, or false alarms hasn't improved the service. For many applications, hybrid operation is the practical answer: conventional coded payloads for authoritative data, semantic updates for prediction and prioritisation.

---

Why AI is central

Modern semantic proposals lean heavily on learned representations. Neural networks extract compact embeddings from images, speech, and language. Generative and foundation models supply contextual reconstruction. Reinforcement learning optimises what to send under changing radio conditions.

AI also moves the implementation boundary. Some inference runs on the device, some at the RAN edge, some in a cloud platform — which makes model placement, hardware acceleration, energy use, and lifecycle management part of the communication system design.

The radio engineer's familiar constraints — link budget, scheduling, reliability, latency — still apply. They now have to be co-optimised with model quality and compute availability, which is a genuinely harder problem than either in isolation.

---

What's actually unsolved

Semantic communication is an active research direction, not a replacement for the reliable packet networks deployed today. Demonstrations are strongest where the source distribution and task are tightly constrained. General-purpose, interoperable semantic networking remains early.

The open questions are substantial:

Metrics. What common measure proves the receiver got the right meaning for a given task? BER is universal and comparable across systems. Semantic fidelity is task-specific and, so far, not.

Robustness. How does the system behave under channel errors, unfamiliar data, adversarial inputs, and model mismatch? A conventional system degrades predictably; a semantic system can fail confidently.

Trust. Can the receiver quantify uncertainty, explain a result, and request a conventional fallback when needed?

Interoperability. How are models versioned, negotiated, updated, and governed across vendors and operators? This is arguably the hardest one — the entire value of a shared knowledge base evaporates if two vendors' models disagree.

Privacy and security. A compact embedding may still reveal sensitive information, and model supply chains need protection of their own.

For 6G, the credible near-term role is as a configurable capability around AI-enabled services — not a claim that every packet should become semantic. The designs likely to win will make the task boundary explicit, retain reliable fallbacks, and demonstrate measurable benefit at the application outcome rather than at the link layer.

---

Takeaway

Semantic communication optimises for task success, not bit fidelity.

The gain is task-selective reduction in constrained paths, not universal compression.

Shared models are both the mechanism and the main risk — mismatch produces confident wrong answers.

Evaluate end to end. Fewer radio bits plus heavy inference is not automatically a win.

Metrics and interoperability are the real blockers, more than encoder design.

It reframes efficiency around useful meaning, and it's most promising where a receiver needs a timely inference rather than a full replica of raw data. Its future depends less on a clever encoder than on shared models, trustworthy evaluation, and disciplined integration with the radio and application stack.

---

Further reading

• AI-native networks — the architectural context this depends on
• Integrated sensing and communication
• ITU-R M.2160 — IMT-2030 framework, AI and communication usage scenario
• Hexa-X-II — European 6G project, semantic and goal-oriented communication work